Files
zenai-ktc-python/counter_live_rknn.py
T

1789 lines
63 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""
Edge production live counter — RTSP + YOLO RKNN + ByteTrack + zone counting.
Runs on RK3588 hardware with RKNN model (320×320 input).
Uses ByteTrack (Kalman filter + two-stage IoU association) for tracking.
Left / right rectangular zones mark the left and right sack feeders.
A sack is counted once when its centroid enters a zone.
"""
import numpy as np
import cv2
import csv
import json
import os
import signal
import socket
import threading
import time
from collections import deque
from datetime import datetime
from pathlib import Path
from dotenv import load_dotenv
load_dotenv()
from rknnlite.api import RKNNLite
from counter_store import CounterStore
# --- config (override via env / .env) ---
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "/opt/jetson-counter")
DB_PATH = os.getenv("DB_PATH", f"{OUTPUT_DIR}/jetson_counter.db")
STATE_FILE = os.getenv("STATE_FILE", f"{OUTPUT_DIR}/current_counter.json")
SOURCE = os.getenv("SOURCE", "rtsp://user:pass@192.168.0.100:554/stream1")
MODEL_PATH = os.getenv("MODEL_PATH", "/opt/jetson-counter/yolo11n.rknn")
CAMERA_NAME = os.getenv("CAMERA_NAME", "CC1")
OBJECT_LABEL = os.getenv("OBJECT_LABEL", "object")
CLASS_OBJECT = os.getenv("CLASS_OBJECT", "object")
# Human-readable feeder names used in count logs, e.g.
# "Total karung di kandang_bawah_feeder_kanan: 5"
FEEDER_LEFT_NAME = os.getenv("FEEDER_LEFT_NAME", f"{CAMERA_NAME}_feeder_kiri")
FEEDER_RIGHT_NAME = os.getenv("FEEDER_RIGHT_NAME", f"{CAMERA_NAME}_feeder_kanan")
DEBUG_TRACKING = os.getenv("DEBUG_TRACKING", "").lower() == "true"
# Left feeder zone (fractions of frame width/height). Absolute pixel overrides
# win when set (ZONE_LEFT_X1 … ZONE_LEFT_Y2).
ZONE_LEFT_X1 = int(os.getenv("ZONE_LEFT_X1")) if os.getenv("ZONE_LEFT_X1") else None
ZONE_LEFT_Y1 = int(os.getenv("ZONE_LEFT_Y1")) if os.getenv("ZONE_LEFT_Y1") else None
ZONE_LEFT_X2 = int(os.getenv("ZONE_LEFT_X2")) if os.getenv("ZONE_LEFT_X2") else None
ZONE_LEFT_Y2 = int(os.getenv("ZONE_LEFT_Y2")) if os.getenv("ZONE_LEFT_Y2") else None
ZONE_LEFT_X1_FRAC = float(os.getenv("ZONE_LEFT_X1_FRAC", "0.00"))
ZONE_LEFT_Y1_FRAC = float(os.getenv("ZONE_LEFT_Y1_FRAC", "0.20"))
ZONE_LEFT_X2_FRAC = float(os.getenv("ZONE_LEFT_X2_FRAC", "0.35"))
ZONE_LEFT_Y2_FRAC = float(os.getenv("ZONE_LEFT_Y2_FRAC", "0.85"))
# Right feeder zone
ZONE_RIGHT_X1 = int(os.getenv("ZONE_RIGHT_X1")) if os.getenv("ZONE_RIGHT_X1") else None
ZONE_RIGHT_Y1 = int(os.getenv("ZONE_RIGHT_Y1")) if os.getenv("ZONE_RIGHT_Y1") else None
ZONE_RIGHT_X2 = int(os.getenv("ZONE_RIGHT_X2")) if os.getenv("ZONE_RIGHT_X2") else None
ZONE_RIGHT_Y2 = int(os.getenv("ZONE_RIGHT_Y2")) if os.getenv("ZONE_RIGHT_Y2") else None
ZONE_RIGHT_X1_FRAC = float(os.getenv("ZONE_RIGHT_X1_FRAC", "0.65"))
ZONE_RIGHT_Y1_FRAC = float(os.getenv("ZONE_RIGHT_Y1_FRAC", "0.20"))
ZONE_RIGHT_X2_FRAC = float(os.getenv("ZONE_RIGHT_X2_FRAC", "1.00"))
ZONE_RIGHT_Y2_FRAC = float(os.getenv("ZONE_RIGHT_Y2_FRAC", "0.85"))
IMGSZ = int(os.getenv("IMGSZ", "320"))
HALF = os.getenv("HALF", "false").lower() == "true"
CONF = float(os.getenv("CONF", "0.3"))
NMS_IOU = float(os.getenv("NMS_IOU", "0.45"))
DEVICE = int(os.getenv("DEVICE", "0"))
# RKNN NPU core mask
CORE_MASK = int(os.getenv("CORE_MASK", "1"))
# YOLO decoder config
NUM_CLASSES = int(os.getenv("NUM_CLASSES", "2"))
SCORE_SIGMOID = os.getenv("SCORE_SIGMOID", "false").lower() == "true"
# ByteTrack settings
TRACK_HIGH_THRESH = float(os.getenv("TRACK_HIGH_THRESH", "0.5"))
TRACK_LOW_THRESH = float(os.getenv("TRACK_LOW_THRESH", "0.1"))
TRACK_MATCH_THRESH = float(os.getenv("TRACK_MATCH_THRESH", "0.8"))
TRACK_BUFFER = int(os.getenv("TRACK_BUFFER", "30"))
TRACK_MIN_HITS = int(os.getenv("TRACK_MIN_HITS", "3"))
# Dedup guard against ID-switch double counts: ignore a second zone entry in the
# same feeder within DEDUP_FRAMES and DEDUP_PX (centroid distance) of a recent count.
DEDUP_FRAMES = int(os.getenv("DEDUP_FRAMES", "15"))
DEDUP_PX = float(os.getenv("DEDUP_PX", "60"))
# Trajectory inheritance across ID switches: when a new track appears, inherit the
# last position of a recently-seen nearby track so zone entry is not missed when
# the ID changes at the zone boundary.
INHERIT_SEC = float(os.getenv("INHERIT_SEC", "1.0"))
INHERIT_PX = float(os.getenv("INHERIT_PX", "60"))
# Per-feeder cooldown: after a count on that feeder, ignore further counts on the
# same feeder for N seconds. 0 = disabled.
ZONE_COOLDOWN_LEFT_SEC = float(os.getenv("ZONE_COOLDOWN_LEFT_SEC", "0"))
ZONE_COOLDOWN_RIGHT_SEC = float(os.getenv("ZONE_COOLDOWN_RIGHT_SEC", "0"))
# Per-feeder dwell: sack must remain inside the zone for N seconds before it is
# counted (and a cross snapshot is saved). 0 = count on first frame in zone.
ZONE_DWELL_LEFT_SEC = float(os.getenv("ZONE_DWELL_LEFT_SEC", "2"))
ZONE_DWELL_RIGHT_SEC = float(os.getenv("ZONE_DWELL_RIGHT_SEC", "2"))
DAILY_CUTOFF_TIME = os.getenv("DAILY_CUTOFF_TIME", "20:00")
EXPORT_CSV = os.getenv("EXPORT_CSV", "true").lower() == "true"
CROSS_CSV = os.getenv("CROSS_CSV", f"{OUTPUT_DIR}/crossings.csv")
# Save an annotated frame snapshot each time a sack finishes its zone dwell and
# the counter increases.
SAVE_CROSS_SNAPSHOT = os.getenv("SAVE_CROSS_SNAPSHOT", "false").lower() == "true"
# Also save one snapshot the first time each object is detected (before it enters
# a zone), named with the same track id so it can be correlated with the count snapshot.
SAVE_DETECT_SNAPSHOT = os.getenv("SAVE_DETECT_SNAPSHOT", "false").lower() == "true"
CROSS_SNAPSHOT_DIR = os.getenv("CROSS_SNAPSHOT_DIR", f"{OUTPUT_DIR}/snapshots")
CROSS_SNAPSHOT_QUALITY = int(os.getenv("CROSS_SNAPSHOT_QUALITY", "85"))
# Retention: delete oldest snapshots when either limit is exceeded (0 = disabled).
CROSS_SNAPSHOT_MAX_FILES = int(os.getenv("CROSS_SNAPSHOT_MAX_FILES", "1000"))
CROSS_SNAPSHOT_MAX_AGE_DAYS = float(os.getenv("CROSS_SNAPSHOT_MAX_AGE_DAYS", "7"))
# Run the cleanup sweep at most every N seconds to limit filesystem scans.
CROSS_SNAPSHOT_CLEANUP_SEC = int(os.getenv("CROSS_SNAPSHOT_CLEANUP_SEC", "3600"))
RATE_WINDOW_SEC = int(os.getenv("RATE_WINDOW_SEC", "60"))
WARMUP_FRAMES = int(os.getenv("WARMUP_FRAMES", "30"))
RECONNECT_DELAY_SEC = int(os.getenv("RECONNECT_DELAY_SEC", "3"))
MAX_RECONNECT_ATTEMPTS = int(os.getenv("MAX_RECONNECT_ATTEMPTS", "0"))
TRACKED_PRUNE_SEC = int(os.getenv("TRACKED_PRUNE_SEC", "300"))
RECORD_VIDEO = os.getenv("RECORD_VIDEO", "false").lower() == "true"
VIDEO_SEGMENT_SEC = int(os.getenv("VIDEO_SEGMENT_SEC", "3600"))
OUTPUT_FPS = int(os.getenv("OUTPUT_FPS", "15"))
LIVE_STREAM_ENABLED = os.getenv("LIVE_STREAM_ENABLED", "false").lower() == "true"
LIVE_STREAM_FRAME_PATH = os.getenv(
"LIVE_STREAM_FRAME_PATH", "/dev/shm/jetson-counter/live_frame.jpg"
)
LIVE_STREAM_QUALITY = int(os.getenv("LIVE_STREAM_QUALITY", "75"))
LIVE_STREAM_EVERY_N = int(os.getenv("LIVE_STREAM_EVERY_N", "2"))
RTSP_FFMPEG_OPTIONS = os.getenv(
"OPENCV_FFMPEG_CAPTURE_OPTIONS",
"rtsp_transport;tcp|fflags;nobuffer|flags;low_delay",
)
IS_LIVE = SOURCE.lower().startswith(("rtsp://", "http://"))
MOTION_DETECTION_ENABLED = os.getenv("MOTION_DETECTION_ENABLED", "false").lower() == "true"
MOTION_THRESHOLD = float(os.getenv("MOTION_THRESHOLD", "5.0"))
# Per-pixel intensity change (0-255) for a pixel to count as "moved".
MOTION_PIXEL_DELTA = int(os.getenv("MOTION_PIXEL_DELTA", "25"))
# Fraction of frame pixels that must change (0-1) to trigger inference. Small,
# so an object entering the edge of the frame is detected immediately.
MOTION_MIN_AREA_FRAC = float(os.getenv("MOTION_MIN_AREA_FRAC", "0.002"))
# Always run inference at least every N frames even if no motion (heartbeat), so a
# slow/stationary object is never missed for long.
MOTION_HEARTBEAT_FRAMES = int(os.getenv("MOTION_HEARTBEAT_FRAMES", "15"))
# --- Runtime control (toggle counting on/off on the fly) ---
# When enabled, the process watches a small JSON control file and honors its
# "counting" flag. Set false to always count (ignore the control file).
CONTROL_ENABLED = os.getenv("CONTROL_ENABLED", "false").lower() == "true"
CONTROL_FILE = os.getenv("CONTROL_FILE", f"{OUTPUT_DIR}/control.json")
# Whether counting is active on startup when no control file exists yet.
CONTROL_DEFAULT_COUNTING = os.getenv("CONTROL_DEFAULT_COUNTING", "true").lower() == "true"
# Re-read the control file at most every N seconds.
CONTROL_POLL_SEC = float(os.getenv("CONTROL_POLL_SEC", "1.0"))
# Optional TCP control socket. When enabled, the counter listens for line-based
# commands so counting can be toggled over the network (in addition to the file).
CONTROL_SOCKET_ENABLED = os.getenv("CONTROL_SOCKET_ENABLED", "false").lower() == "true"
CONTROL_SOCKET_HOST = os.getenv("CONTROL_SOCKET_HOST", "127.0.0.1")
CONTROL_SOCKET_PORT = int(os.getenv("CONTROL_SOCKET_PORT", "5090"))
CROSS_FLASH_FRAMES = 12
POPUP_LIFETIME = 20
ZONE_PULSE_FRAMES = 12
COUNT_PULSE_FRAMES = 15
C_ZONE_LEFT = (80, 220, 100) # green — left feeder
C_ZONE_RIGHT = (0, 165, 255) # orange — right feeder
C_ZONE_FILL_ALPHA = 0.18
SKELETON = [(0, 1), (4, 3), (1, 2), (3, 2), (2, 6), (2, 5), (2, 7), (7, 8)]
SK_COLORS = [
(0, 255, 255),
(0, 255, 255),
(255, 0, 255),
(255, 0, 255),
(0, 255, 0),
(255, 255, 0),
(0, 0, 255),
(200, 200, 0),
]
C_PANEL = (28, 24, 18)
C_BORDER = (90, 85, 75)
C_ACCENT = (255, 200, 60)
C_GREEN = (80, 220, 100)
C_TEXT = (235, 235, 235)
C_MUTED = (150, 150, 150)
C_OBJECT_BOX = (0, 165, 255)
shutdown_requested = False
def request_shutdown(signum, frame):
global shutdown_requested
shutdown_requested = True
print("\nShutdown requested — finishing current frame...")
signal.signal(signal.SIGINT, request_shutdown)
signal.signal(signal.SIGTERM, request_shutdown)
# =============================================================================
# YOLO output decoder (NMS only — boxes are pre-decoded by the model)
# =============================================================================
def _nms(boxes, scores, iou_thr=0.45):
order = np.argsort(scores)[::-1]
keep = []
while len(order) > 0:
idx = order[0]
keep.append(idx)
if len(order) == 1:
break
xx1 = np.maximum(boxes[idx, 0], boxes[order[1:], 0])
yy1 = np.maximum(boxes[idx, 1], boxes[order[1:], 1])
xx2 = np.minimum(boxes[idx, 2], boxes[order[1:], 2])
yy2 = np.minimum(boxes[idx, 3], boxes[order[1:], 3])
w = np.maximum(0.0, xx2 - xx1)
h = np.maximum(0.0, yy2 - yy1)
inter = w * h
area_i = (boxes[idx, 2] - boxes[idx, 0]) * (boxes[idx, 3] - boxes[idx, 1])
area_o = (boxes[order[1:], 2] - boxes[order[1:], 0]) * (
boxes[order[1:], 3] - boxes[order[1:], 1]
)
iou = inter / (area_i + area_o - inter + 1e-16)
order = order[1:][iou < iou_thr]
return np.array(keep)
# =============================================================================
# IoU helpers (xyxy format)
# =============================================================================
def _ious_xyxy(boxes_a, boxes_b):
"""Pairwise IoU: (N,4) vs (M,4) → (N,M) matrix."""
n, m = len(boxes_a), len(boxes_b)
if n == 0 or m == 0:
return np.zeros((n, m), dtype=np.float32)
xx1 = np.maximum(boxes_a[:, None, 0], boxes_b[None, :, 0])
yy1 = np.maximum(boxes_a[:, None, 1], boxes_b[None, :, 1])
xx2 = np.minimum(boxes_a[:, None, 2], boxes_b[None, :, 2])
yy2 = np.minimum(boxes_a[:, None, 3], boxes_b[None, :, 3])
iw = np.maximum(0.0, xx2 - xx1)
ih = np.maximum(0.0, yy2 - yy1)
inter = iw * ih
area_a = (boxes_a[:, 2] - boxes_a[:, 0]) * (boxes_a[:, 3] - boxes_a[:, 1])
area_b = (boxes_b[:, 2] - boxes_b[:, 0]) * (boxes_b[:, 3] - boxes_b[:, 1])
return inter / (area_a[:, None] + area_b[None, :] - inter + 1e-16)
def _greedy_match(cost_matrix, threshold=0.3):
"""Greedy linear assignment. Returns pairs (row_idx, col_idx)."""
if cost_matrix.size == 0:
return []
n, m = cost_matrix.shape
flat = [(cost_matrix[i, j], i, j) for i in range(n) for j in range(m)]
flat.sort()
row_used = set()
col_used = set()
pairs = []
for cost, i, j in flat:
if cost >= threshold:
break
if i in row_used or j in col_used:
continue
row_used.add(i)
col_used.add(j)
pairs.append((i, j))
return pairs
# =============================================================================
# Kalman filter box tracker (state: x, y, w, h, vx, vy, vw, vh)
# =============================================================================
class KalmanBoxTracker:
count = 0
def __init__(self, bbox_xyxy):
KalmanBoxTracker.count += 1
self.track_id = KalmanBoxTracker.count
x1, y1, x2, y2 = bbox_xyxy
w, h = x2 - x1, y2 - y1
x, y = x1 + w / 2, y1 + h / 2
self.kf = _KalmanFilter()
self.kf.x[:4, 0] = np.array([x, y, w, h], dtype=np.float32)
self.time_since_update = 0
self.hits = 1
self.hit_streak = 1
self.age = 1
def predict(self):
if self.kf.x[6] + self.kf.x[2] <= 0:
self.kf.x[6] *= 0.0
self.kf.predict()
self.age += 1
self.time_since_update += 1
def update(self, bbox_xyxy):
self.time_since_update = 0
self.hits += 1
self.hit_streak += 1
x1, y1, x2, y2 = bbox_xyxy
w, h = x2 - x1, y2 - y1
x, y = x1 + w / 2, y1 + h / 2
self.kf.update(np.array([x, y, w, h], dtype=np.float32))
def get_state(self):
"""Returns xyxy bbox from Kalman state."""
xx = self.kf.x[:4, 0]
x, y, w, h = xx[0], xx[1], xx[2], xx[3]
x1 = x - w / 2
y1 = y - h / 2
x2 = x + w / 2
y2 = y + h / 2
return np.array([x1, y1, x2, y2], dtype=np.float32)
def get_cx(self):
return float(self.kf.x[0, 0])
def get_cy(self):
return float(self.kf.x[1, 0])
class _KalmanFilter:
"""8-state constant-velocity Kalman filter for bounding box tracking."""
def __init__(self):
ndim, dt = 4, 1.0
self.motion_mat = np.eye(2 * ndim, 2 * ndim, dtype=np.float32)
for i in range(ndim):
self.motion_mat[i, ndim + i] = dt
self.update_mat = np.eye(ndim, 2 * ndim, dtype=np.float32)
self._std_weight_position = 1.0 / 20
self._std_weight_velocity = 1.0 / 160
self.x = np.zeros((8, 1), dtype=np.float32)
self.P = np.eye(8, dtype=np.float32) * 10.0
def predict(self):
std_pos = [
self._std_weight_position * self.x[2],
self._std_weight_position * self.x[3],
self._std_weight_position * self.x[2],
self._std_weight_position * self.x[3],
]
std_vel = [
self._std_weight_velocity * self.x[2],
self._std_weight_velocity * self.x[3],
self._std_weight_velocity * self.x[2],
self._std_weight_velocity * self.x[3],
]
Q = np.diag(np.square(np.concatenate([std_pos, std_vel])))
self.x = self.motion_mat @ self.x
self.P = self.motion_mat @ self.P @ self.motion_mat.T + Q
def update(self, z):
R = np.diag(
np.square(
[
self._std_weight_position * z[2],
self._std_weight_position * z[3],
self._std_weight_position * z[2],
self._std_weight_position * z[3],
]
)
)
H = self.update_mat
S = H @ self.P @ H.T + R
K = self.P @ H.T @ np.linalg.inv(S)
y = z.reshape(4, 1) - H @ self.x
self.x = self.x + K @ y
I_KH = np.eye(8) - K @ H
self.P = I_KH @ self.P @ I_KH.T + K @ R @ K.T
# =============================================================================
# ByteTrack multi-object tracker
# =============================================================================
class ByteTracker:
"""ByteTrack: two-stage association with Kalman filter prediction."""
def __init__(
self,
track_high_thresh=0.5,
track_low_thresh=0.1,
match_thresh=0.8,
track_buffer=30,
min_hits=3,
):
self.high_thresh = track_high_thresh
self.low_thresh = track_low_thresh
self.match_thresh = match_thresh
self.track_buffer = track_buffer
self.min_hits = min_hits
self.tracked_tracks = []
self.lost_tracks = []
self.removed_tracks = []
self.frame_id = 0
def update(self, boxes_xyxy, scores):
self.frame_id += 1
# --- separate detections by score ---
if len(boxes_xyxy) > 0:
remain = scores > self.low_thresh
remain_orig_idx = np.where(remain)[0]
dets = boxes_xyxy[remain]
det_scores = scores[remain]
is_high = det_scores > self.high_thresh
is_low = ~is_high
else:
remain_orig_idx = np.zeros(0, dtype=np.int64)
dets = np.zeros((0, 4), dtype=np.float32)
det_scores = np.zeros(0, dtype=np.float32)
is_high = np.zeros(0, dtype=bool)
is_low = np.zeros(0, dtype=bool)
# --- Kalman predict all existing tracks ---
track_pool = self.tracked_tracks + self.lost_tracks
num_tracks = len(track_pool)
# Per-frame tracking results
matched_track_idx = set()
det_to_track = {}
tracked_map = {}
lost_map = {}
# Pre-allocate these for scoping
high_idx = np.array([], dtype=np.int64)
low_idx = np.array([], dtype=np.int64)
match_pairs_high = []
if num_tracks > 0:
track_boxes = np.zeros((num_tracks, 4), dtype=np.float32)
for ti, trk in enumerate(track_pool):
trk.predict()
track_boxes[ti] = trk.get_state()
# --- first association: high-score ↔ all tracks ---
high_idx = np.where(is_high)[0]
high_dets = dets[is_high]
unmatched_tracks = list(range(num_tracks))
if len(high_dets) > 0:
iou_mat = _ious_xyxy(high_dets, track_boxes)
cost_mat = 1.0 - iou_mat
matches = _greedy_match(cost_mat, threshold=1.0 - self.match_thresh)
for di, ti in matches:
det_global = int(high_idx[di])
orig_idx = int(remain_orig_idx[det_global])
track_pool[ti].update(dets[det_global])
track_pool[ti].hit_streak = max(1, track_pool[ti].hit_streak)
matched_track_idx.add(ti)
det_to_track[orig_idx] = track_pool[ti].track_id
tracked_map[track_pool[ti].track_id] = (track_pool[ti].get_cx(), track_pool[ti].get_cy())
match_pairs_high.append((det_global, ti))
unmatched_tracks = [
t for t in range(num_tracks) if t not in matched_track_idx
]
# --- second association: low-score ↔ unmatched tracks ---
low_idx = np.where(is_low)[0]
low_dets = dets[is_low]
if len(low_dets) > 0 and len(unmatched_tracks) > 0:
unmatched_boxes = track_boxes[unmatched_tracks]
iou_mat = _ious_xyxy(low_dets, unmatched_boxes)
cost_mat = 1.0 - iou_mat
matches2 = _greedy_match(
cost_mat, threshold=1.0 - self.match_thresh
)
for di, uti in matches2:
det_global = int(low_idx[di])
pool_idx = unmatched_tracks[uti]
orig_idx = int(remain_orig_idx[det_global])
track_pool[pool_idx].update(dets[det_global])
track_pool[pool_idx].hit_streak = max(
1, track_pool[pool_idx].hit_streak
)
matched_track_idx.add(pool_idx)
det_to_track[orig_idx] = track_pool[pool_idx].track_id
tracked_map[track_pool[pool_idx].track_id] = (
track_pool[pool_idx].get_cx(),
track_pool[pool_idx].get_cy(),
)
# --- reset hit_streak for unmatched tracks ---
for ti, trk in enumerate(track_pool):
if ti not in matched_track_idx:
trk.hit_streak = 0
# --- lifecycle management ---
new_tracked = []
new_lost = []
for trk in track_pool:
if trk.time_since_update > self.track_buffer:
self.removed_tracks.append(trk)
elif trk.time_since_update > 0:
new_lost.append(trk)
else:
new_tracked.append(trk)
self.tracked_tracks = new_tracked
self.lost_tracks = new_lost
# --- confirmed tracks (both tracked and lost) ---
for trk in self.tracked_tracks + self.lost_tracks:
if trk.hit_streak >= self.min_hits or trk.hits >= self.min_hits:
tracked_map.setdefault(trk.track_id, (trk.get_cx(), trk.get_cy()))
for trk in self.lost_tracks:
if trk.hit_streak >= self.min_hits or trk.hits >= self.min_hits:
lost_map[trk.track_id] = (trk.get_cx(), trk.get_cy())
# --- new tracks from unmatched high-score dets ---
high_all = np.where(is_high)[0]
matched_det_ids = set(det_to_track.keys())
for dg in high_all:
orig_idx = int(remain_orig_idx[int(dg)])
if orig_idx not in matched_det_ids:
trk = KalmanBoxTracker(dets[int(dg)])
self.tracked_tracks.append(trk)
det_to_track[orig_idx] = trk.track_id
tracked_map[trk.track_id] = (trk.get_cx(), trk.get_cy())
return tracked_map, det_to_track, lost_map
# =============================================================================
# RKNN YOLO wrapper (detect output format: (1, 4+num_classes, N))
# =============================================================================
class RKNNYOLO:
def __init__(
self,
model_path,
core_mask=1,
imgsz=320,
conf=0.3,
iou=0.45,
num_classes=2,
num_keypoints=0,
score_sigmoid=False,
):
self.imgsz = imgsz
self.conf = conf
self.iou = iou
self.num_classes = num_classes
self.num_keypoints = num_keypoints
self.score_sigmoid = score_sigmoid
self.rknn = RKNNLite(verbose=False)
ret = self.rknn.load_rknn(model_path)
if ret != 0:
raise RuntimeError(f"Failed to load RKNN model: {model_path}")
ret = self.rknn.init_runtime(core_mask=core_mask)
if ret != 0:
raise RuntimeError(f"Failed to init RKNN runtime (core_mask={core_mask})")
try:
sdk_ver = self.rknn.get_sdk_version()
print(f"RKNN SDK version: {sdk_ver}")
except Exception:
pass
print(f"RKNN model loaded: {model_path} imgsz={imgsz} core_mask={core_mask}")
def _preprocess(self, frame):
h0, w0 = frame.shape[:2]
scale = min(self.imgsz / h0, self.imgsz / w0)
nh, nw = int(h0 * scale), int(w0 * scale)
resized = cv2.resize(frame, (nw, nh), interpolation=cv2.INTER_LINEAR)
letterbox = np.full((self.imgsz, self.imgsz, 3), 114, dtype=np.uint8)
dy = (self.imgsz - nh) // 2
dx = (self.imgsz - nw) // 2
letterbox[dy : dy + nh, dx : dx + nw] = resized
rgb = cv2.cvtColor(letterbox, cv2.COLOR_BGR2RGB)
gains = np.array([scale, scale, dy, dx], dtype=np.float32)
return rgb, gains
def __call__(self, frame):
h0, w0 = frame.shape[:2]
rgb, gains = self._preprocess(frame)
scale, _, pad_y, pad_x = gains
inp = np.expand_dims(rgb, axis=0)
inp = np.ascontiguousarray(inp.astype(np.uint8))
outputs = self.rknn.inference(inputs=[inp])
if len(outputs) == 0:
return []
out = outputs[0]
out = np.squeeze(out, axis=0)
if out.shape[0] == self.num_classes + 4:
out = out.T
boxes_cxcywh = out[:, :4].copy()
cls_raw = out[:, 4:].copy()
if self.score_sigmoid:
cls_scores = 1.0 / (1.0 + np.exp(-np.clip(cls_raw, -10, 10)))
else:
cls_scores = cls_raw
boxes_xyxy = np.stack(
[
boxes_cxcywh[:, 0] - boxes_cxcywh[:, 2] / 2,
boxes_cxcywh[:, 1] - boxes_cxcywh[:, 3] / 2,
boxes_cxcywh[:, 0] + boxes_cxcywh[:, 2] / 2,
boxes_cxcywh[:, 1] + boxes_cxcywh[:, 3] / 2,
],
axis=1,
)
max_scores = cls_scores.max(axis=1)
class_ids = cls_scores.argmax(axis=1)
mask = max_scores > self.conf
if mask.sum() == 0:
return []
bboxes = boxes_xyxy[mask].astype(np.float32)
scores = max_scores[mask].astype(np.float32)
clses = class_ids[mask]
bboxes[:, 0] = (bboxes[:, 0] - pad_x) / scale
bboxes[:, 1] = (bboxes[:, 1] - pad_y) / scale
bboxes[:, 2] = (bboxes[:, 2] - pad_x) / scale
bboxes[:, 3] = (bboxes[:, 3] - pad_y) / scale
bboxes[:, 0] = np.clip(bboxes[:, 0], 0, w0)
bboxes[:, 1] = np.clip(bboxes[:, 1], 0, h0)
bboxes[:, 2] = np.clip(bboxes[:, 2], 0, w0)
bboxes[:, 3] = np.clip(bboxes[:, 3], 0, h0)
detections = []
for cls_id in range(self.num_classes):
idx = np.where(clses == cls_id)[0]
if len(idx) == 0:
continue
keep = _nms(bboxes[idx], scores[idx], iou_thr=self.iou)
for k in keep:
j = idx[k]
detections.append(
{
"bbox": bboxes[j].tolist(),
"score": float(scores[j]),
"cls": int(clses[j]),
"keypoints": None,
}
)
return detections
def release(self):
self.rknn.release()
# =============================================================================
# Drawing helpers
# =============================================================================
def _zone_coord(abs_val, frac, span):
if abs_val is not None:
return int(abs_val)
return int(span * frac)
def resolve_zone_left(frame_width, frame_height):
"""Return (x1, y1, x2, y2) for the left feeder zone."""
x1 = _zone_coord(ZONE_LEFT_X1, ZONE_LEFT_X1_FRAC, frame_width)
y1 = _zone_coord(ZONE_LEFT_Y1, ZONE_LEFT_Y1_FRAC, frame_height)
x2 = _zone_coord(ZONE_LEFT_X2, ZONE_LEFT_X2_FRAC, frame_width)
y2 = _zone_coord(ZONE_LEFT_Y2, ZONE_LEFT_Y2_FRAC, frame_height)
return (min(x1, x2), min(y1, y2), max(x1, x2), max(y1, y2))
def resolve_zone_right(frame_width, frame_height):
"""Return (x1, y1, x2, y2) for the right feeder zone."""
x1 = _zone_coord(ZONE_RIGHT_X1, ZONE_RIGHT_X1_FRAC, frame_width)
y1 = _zone_coord(ZONE_RIGHT_Y1, ZONE_RIGHT_Y1_FRAC, frame_height)
x2 = _zone_coord(ZONE_RIGHT_X2, ZONE_RIGHT_X2_FRAC, frame_width)
y2 = _zone_coord(ZONE_RIGHT_Y2, ZONE_RIGHT_Y2_FRAC, frame_height)
return (min(x1, x2), min(y1, y2), max(x1, x2), max(y1, y2))
def point_in_zone(cx, cy, zone):
x1, y1, x2, y2 = zone
return x1 <= cx <= x2 and y1 <= cy <= y2
def is_duplicate_cross(recent, cx, cy, frame_idx):
"""True if a count near (cx, cy) happened within the dedup window (ID-switch guard)."""
while recent and frame_idx - recent[0][0] > DEDUP_FRAMES:
recent.popleft()
for _, prev_cx, prev_cy in recent:
if abs(prev_cx - cx) <= DEDUP_PX and abs(prev_cy - cy) <= DEDUP_PX:
return True
return False
def feeder_cooldown_sec(feeder):
if feeder == "left":
return ZONE_COOLDOWN_LEFT_SEC
return ZONE_COOLDOWN_RIGHT_SEC
def feeder_dwell_sec(feeder):
if feeder == "left":
return ZONE_DWELL_LEFT_SEC
return ZONE_DWELL_RIGHT_SEC
def resolve_feeder_at(cx, cy, zone_left, zone_right):
"""Return 'left' | 'right' | None for the zone containing (cx, cy)."""
now_left = point_in_zone(cx, cy, zone_left)
now_right = point_in_zone(cx, cy, zone_right)
if now_left and now_right:
lx1, ly1, lx2, ly2 = zone_left
rx1, ry1, rx2, ry2 = zone_right
dl = abs(cx - (lx1 + lx2) / 2) + abs(cy - (ly1 + ly2) / 2)
dr = abs(cx - (rx1 + rx2) / 2) + abs(cy - (ry1 + ry2) / 2)
return "left" if dl <= dr else "right"
if now_left:
return "left"
if now_right:
return "right"
return None
def feeder_display_name(feeder):
if feeder == "left":
return FEEDER_LEFT_NAME
return FEEDER_RIGHT_NAME
def feeder_in_cooldown(feeder, mono, last_count_mono):
"""True if this feeder is still inside its post-count cooldown window."""
cooldown = feeder_cooldown_sec(feeder)
if cooldown <= 0:
return False
last = last_count_mono.get(feeder)
if last is None:
return False
return (mono - last) < cooldown
def _inherit_prev(tracked, new_tid, cx, cy, mono, max_age, max_px):
"""Find a recently-seen nearby track for ID-switch continuation.
Returns (cx, cy, mono, source_tid) or None.
"""
best = None
best_dist = max_px
for tid, (tcx, tcy, ts) in tracked.items():
if tid == new_tid:
continue
if mono - ts > max_age:
continue
if abs(tcy - cy) > max_px:
continue
dist = abs(tcx - cx)
if dist <= best_dist:
best_dist = dist
best = (tcx, tcy, ts, tid)
return best
def _default_zone_state():
"""Per-track feeder-zone state.
counted: already contributed one LEFT or RIGHT count (at most one per object).
side: 'left' | 'right' | None after a count.
dwell_feeder / dwell_since: which zone the track is currently dwelling in,
and when continuous presence in that zone began.
"""
return {
"counted": False,
"side": None,
"dwell_feeder": None,
"dwell_since": None,
}
def _copy_zone_state(src):
return {
"counted": bool(src.get("counted", False)),
"side": src.get("side"),
"dwell_feeder": src.get("dwell_feeder"),
"dwell_since": src.get("dwell_since"),
}
def now_str():
return datetime.now().strftime("%Y-%m-%d %H:%M:%S")
def read_counting_flag(default=True):
"""Read the 'counting' flag from the control file. Returns default on any error."""
try:
with open(CONTROL_FILE, "r", encoding="utf-8") as f:
data = json.load(f)
return bool(data.get("counting", default))
except FileNotFoundError:
return default
except Exception:
return default
def write_control_file(counting):
"""Create/update the control file atomically (used to seed defaults)."""
try:
Path(CONTROL_FILE).parent.mkdir(parents=True, exist_ok=True)
tmp = f"{CONTROL_FILE}.tmp"
with open(tmp, "w", encoding="utf-8") as f:
json.dump({"counting": bool(counting)}, f)
os.replace(tmp, CONTROL_FILE)
except Exception as exc:
print(f"[{now_str()}] Failed to write control file: {exc}")
def start_control_socket():
"""Start a TCP server for runtime control. Commands (newline-terminated):
START | RESUME | ON -> counting on
STOP | PAUSE | OFF -> counting off
TOGGLE -> flip
STATUS | GET -> report current state
It writes the shared control file, so the main loop's file-poll applies it.
Returns the server socket (call .close() to stop)."""
srv = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
srv.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
srv.bind((CONTROL_SOCKET_HOST, CONTROL_SOCKET_PORT))
srv.listen(5)
def handle(conn, addr):
with conn:
conn.settimeout(30)
try:
buf = b""
while not shutdown_requested:
try:
chunk = conn.recv(256)
except socket.timeout:
break
if not chunk:
break
buf += chunk
while b"\n" in buf:
line, buf = buf.split(b"\n", 1)
cmd = line.decode("utf-8", "ignore").strip().upper()
if not cmd:
continue
current = read_counting_flag(CONTROL_DEFAULT_COUNTING)
if cmd in ("START", "RESUME", "ON"):
write_control_file(True)
resp = "OK counting=on"
elif cmd in ("STOP", "PAUSE", "OFF"):
write_control_file(False)
resp = "OK counting=off"
elif cmd == "TOGGLE":
write_control_file(not current)
resp = f"OK counting={'off' if current else 'on'}"
elif cmd in ("STATUS", "GET"):
resp = f"OK counting={'on' if current else 'off'}"
else:
resp = "ERR unknown command"
conn.sendall((resp + "\n").encode("utf-8"))
except Exception:
pass
def loop():
print(f"Control socket listening on {CONTROL_SOCKET_HOST}:{CONTROL_SOCKET_PORT}")
while not shutdown_requested:
try:
conn, addr = srv.accept()
except OSError:
break
t = threading.Thread(target=handle, args=(conn, addr), daemon=True)
t.start()
threading.Thread(target=loop, daemon=True).start()
return srv
def open_capture(source):
if source.lower().startswith(("rtsp://", "http://")):
os.environ["OPENCV_FFMPEG_CAPTURE_OPTIONS"] = RTSP_FFMPEG_OPTIONS
cap = cv2.VideoCapture(source, cv2.CAP_FFMPEG)
cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
return cap
def warmup_stream(cap, n=WARMUP_FRAMES):
print("Warming up stream...")
for _ in range(n):
cap.read()
print("Stream ready!")
def open_video_writer(path, w, h, fps):
return cv2.VideoWriter(path, cv2.VideoWriter_fourcc(*"avc1"), fps, (w, h))
class CsvLogger:
def __init__(self, path, header):
Path(path).parent.mkdir(parents=True, exist_ok=True)
new_file = not Path(path).exists() or Path(path).stat().st_size == 0
self.file = open(path, "a", newline="", buffering=1)
self.writer = csv.writer(self.file)
if new_file:
self.writer.writerow(header)
self.file.flush()
def write_row(self, row):
self.writer.writerow(row)
self.file.flush()
def close(self):
self.file.close()
class VideoSegmentWriter:
def __init__(self, output_dir, w, h, fps, segment_sec):
self.output_dir = Path(output_dir)
self.output_dir.mkdir(parents=True, exist_ok=True)
self.w, self.h, self.fps = w, h, fps
self.segment_sec = segment_sec
self.segment_start = time.monotonic()
self.writer = None
self._open_next()
def _segment_path(self):
ts = datetime.now().strftime("%Y%m%d_%H%M%S")
return str(self.output_dir / f"live_{ts}.mp4")
def _open_next(self):
if self.writer is not None:
self.writer.release()
path = self._segment_path()
self.writer = open_video_writer(path, self.w, self.h, self.fps)
self.segment_start = time.monotonic()
print(f"Recording segment: {path}")
def write(self, frame):
if time.monotonic() - self.segment_start >= self.segment_sec:
self._open_next()
self.writer.write(frame)
def release(self):
if self.writer is not None:
self.writer.release()
def prune_stale_tracks(tracked, now_mono):
stale = [
tid for tid, (_, _, ts) in tracked.items() if now_mono - ts > TRACKED_PRUNE_SEC
]
for tid in stale:
del tracked[tid]
def cleanup_snapshots(snapshot_dir, max_files, max_age_days):
"""Delete oldest / expired crossing snapshots to bound disk usage."""
d = Path(snapshot_dir)
if not d.is_dir():
return
files = sorted(d.rglob("*.jpg"), key=lambda p: p.stat().st_mtime)
if max_age_days > 0:
cutoff = time.time() - max_age_days * 86400
for p in list(files):
if p.stat().st_mtime < cutoff:
p.unlink(missing_ok=True)
files.remove(p)
if max_files > 0 and len(files) > max_files:
for p in files[: len(files) - max_files]:
p.unlink(missing_ok=True)
def overlay_rect(img, x1, y1, x2, y2, color, alpha=0.65):
x1, y1 = max(0, x1), max(0, y1)
x2, y2 = min(img.shape[1], x2), min(img.shape[0], y2)
if x2 <= x1 or y2 <= y1:
return
roi = img[y1:y2, x1:x2]
patch = np.full_like(roi, color, dtype=np.uint8)
cv2.addWeighted(patch, alpha, roi, 1 - alpha, 0, roi)
def draw_pill(img, text, x, y, bg, fg=C_TEXT, font_scale=0.45, pad_x=6, pad_y=4):
font = cv2.FONT_HERSHEY_SIMPLEX
(tw, th), baseline = cv2.getTextSize(text, font, font_scale, 1)
x1, y1 = x, y - th - pad_y
x2, y2 = x + tw + pad_x * 2, y + baseline + pad_y
cv2.rectangle(img, (x1, y1), (x2, y2), bg, -1)
cv2.rectangle(img, (x1, y1), (x2, y2), C_BORDER, 1)
cv2.putText(img, text, (x + pad_x, y), font, font_scale, fg, 1, cv2.LINE_AA)
def draw_elegant_zone(img, zone, pulse_remaining=0, label="ZONE", color=C_ZONE_LEFT):
x1, y1, x2, y2 = zone
strength = pulse_remaining / max(ZONE_PULSE_FRAMES, 1)
fill_alpha = C_ZONE_FILL_ALPHA + 0.12 * strength
overlay_rect(img, x1, y1, x2, y2, color, alpha=fill_alpha)
thickness = 2 + int(2 * strength)
cv2.rectangle(img, (x1, y1), (x2, y2), color, thickness, cv2.LINE_AA)
# Corner ticks
tick = max(12, min(x2 - x1, y2 - y1) // 8)
for (ax, ay, dx, dy) in (
(x1, y1, 1, 1),
(x2, y1, -1, 1),
(x1, y2, 1, -1),
(x2, y2, -1, -1),
):
cv2.line(img, (ax, ay), (ax + dx * tick, ay), color, 2, cv2.LINE_AA)
cv2.line(img, (ax, ay), (ax, ay + dy * tick), color, 2, cv2.LINE_AA)
cv2.putText(
img,
label,
(x1 + 10, y1 + 22),
cv2.FONT_HERSHEY_SIMPLEX,
0.5,
color,
1,
cv2.LINE_AA,
)
def draw_zone_count(img, zone, count, label, color, pulse_remaining=0):
text = str(count)
font = cv2.FONT_HERSHEY_SIMPLEX
boost = 0.35 * (pulse_remaining / max(COUNT_PULSE_FRAMES, 1))
font_scale, thickness = 1.2 + boost, 3
(tw, th), _ = cv2.getTextSize(text, font, font_scale, thickness)
(lw, lh), _ = cv2.getTextSize(label, font, 0.45, 1)
pad = 12
box_w = max(tw, lw) + pad * 2
box_h = th + lh + pad * 2 + 6
x1, y1, x2, y2 = zone
bx1 = x1 + ((x2 - x1) - box_w) // 2
by1 = max(8, y1 - box_h - 8)
bx2 = bx1 + box_w
by2 = by1 + box_h
overlay_rect(img, bx1, by1, bx2, by2, C_PANEL, alpha=0.78)
cv2.rectangle(img, (bx1, by1), (bx2, by2), color, 2)
tx = bx1 + (box_w - tw) // 2
ty = by1 + pad + th
cv2.putText(img, text, (tx, ty), font, font_scale, color, thickness, cv2.LINE_AA)
slx = bx1 + (box_w - lw) // 2
sly = ty + lh + 6
cv2.putText(img, label, (slx, sly), font, 0.45, C_MUTED, 1, cv2.LINE_AA)
def draw_hud(img, w, total, total_left, total_right, elapsed_sec, rate):
bar_h = 40
overlay_rect(img, 0, 0, w, bar_h, C_PANEL, alpha=0.72)
cv2.line(img, (0, bar_h), (w, bar_h), C_BORDER, 1)
cv2.putText(
img, "TOTAL", (14, 14), cv2.FONT_HERSHEY_SIMPLEX, 0.32, C_MUTED, 1, cv2.LINE_AA
)
cv2.putText(
img,
str(total),
(14, 32),
cv2.FONT_HERSHEY_SIMPLEX,
0.55,
C_GREEN,
1,
cv2.LINE_AA,
)
cv2.putText(
img, "LEFT", (100, 14), cv2.FONT_HERSHEY_SIMPLEX, 0.32, C_MUTED, 1, cv2.LINE_AA
)
cv2.putText(
img,
str(total_left),
(100, 32),
cv2.FONT_HERSHEY_SIMPLEX,
0.55,
C_ZONE_LEFT,
1,
cv2.LINE_AA,
)
cv2.putText(
img, "RIGHT", (180, 14), cv2.FONT_HERSHEY_SIMPLEX, 0.32, C_MUTED, 1, cv2.LINE_AA
)
cv2.putText(
img,
str(total_right),
(180, 32),
cv2.FONT_HERSHEY_SIMPLEX,
0.55,
C_ZONE_RIGHT,
1,
cv2.LINE_AA,
)
cv2.putText(
img,
"UPTIME",
(270, 14),
cv2.FONT_HERSHEY_SIMPLEX,
0.32,
C_MUTED,
1,
cv2.LINE_AA,
)
cv2.putText(
img,
f"{elapsed_sec / 3600:.1f}h",
(270, 32),
cv2.FONT_HERSHEY_SIMPLEX,
0.45,
C_TEXT,
1,
cv2.LINE_AA,
)
cv2.putText(
img, "RATE", (360, 14), cv2.FONT_HERSHEY_SIMPLEX, 0.32, C_MUTED, 1, cv2.LINE_AA
)
cv2.putText(
img,
f"{rate:.1f}/min",
(360, 32),
cv2.FONT_HERSHEY_SIMPLEX,
0.45,
C_ACCENT,
1,
cv2.LINE_AA,
)
def draw_footer(img, w, h, frame_idx, live_tag, inf_ms=0.0, model_name=""):
bar_h = 28
overlay_rect(img, 0, h - bar_h, w, h, C_PANEL, alpha=0.55)
cv2.putText(
img,
f"{live_tag} | {model_name} | Frame {frame_idx} | Inf {inf_ms:.1f}ms",
(12, h - 9),
cv2.FONT_HERSHEY_SIMPLEX,
0.45,
C_MUTED,
1,
cv2.LINE_AA,
)
def draw_skeleton_bold(img, kpts):
for (a, b), color in zip(SKELETON, SK_COLORS):
if a < len(kpts) and b < len(kpts):
xa, ya = int(kpts[a][0]), int(kpts[a][1])
xb, yb = int(kpts[b][0]), int(kpts[b][1])
if xa > 0 and ya > 0 and xb > 0 and yb > 0:
cv2.line(img, (xa, ya), (xb, yb), color, 3, cv2.LINE_AA)
for kp in kpts:
x, y = int(kp[0]), int(kp[1])
if x > 0 and y > 0:
cv2.circle(img, (x, y), 6, (255, 255, 255), -1, cv2.LINE_AA)
cv2.circle(img, (x, y), 6, (40, 40, 40), 2, cv2.LINE_AA)
def draw_popups(img, popups, frame_idx):
alive = []
for pop in popups:
age = frame_idx - pop["born"]
if age > POPUP_LIFETIME:
continue
alive.append(pop)
fade = 1.0 - age / POPUP_LIFETIME
y = pop["y"] - int(age * 1.8)
color = (int(C_GREEN[0] * fade), int(C_GREEN[1] * fade), int(C_GREEN[2] * fade))
cv2.putText(
img,
pop["text"],
(pop["x"], y),
cv2.FONT_HERSHEY_SIMPLEX,
0.7,
color,
2,
cv2.LINE_AA,
)
return alive
def connect_stream(source, warmup=WARMUP_FRAMES):
attempts = 0
while not shutdown_requested:
cap = open_capture(source)
if not cap.isOpened():
attempts += 1
if MAX_RECONNECT_ATTEMPTS and attempts >= MAX_RECONNECT_ATTEMPTS:
raise RuntimeError(
f"Cannot open source after {attempts} attempts: {source}"
)
print(f"Cannot open source, retry in {RECONNECT_DELAY_SEC}s...")
time.sleep(RECONNECT_DELAY_SEC)
continue
if warmup > 0 and source.lower().startswith(("rtsp://", "http://")):
warmup_stream(cap, warmup)
w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
fps = cap.get(cv2.CAP_PROP_FPS)
if not fps or fps <= 1:
fps = OUTPUT_FPS
return cap, w, h, fps
return None, 0, 0, OUTPUT_FPS
# =============================================================================
# Main loop
# =============================================================================
def run():
global shutdown_requested
store = CounterStore(
db_path=DB_PATH,
state_file=STATE_FILE,
camera_name=CAMERA_NAME,
object_label=OBJECT_LABEL,
cutoff_time=DAILY_CUTOFF_TIME,
logger=lambda msg: print(f"[{now_str()}] {msg}"),
)
store.start_cutoff_watcher()
cross_logger = None
if EXPORT_CSV:
cross_logger = CsvLogger(
CROSS_CSV, ["counting_date", "frame", "feeder", "object_id"]
)
model = RKNNYOLO(
model_path=MODEL_PATH,
core_mask=CORE_MASK,
imgsz=IMGSZ,
conf=CONF,
iou=NMS_IOU,
num_classes=NUM_CLASSES,
score_sigmoid=SCORE_SIGMOID,
)
object_cls = int(os.getenv("OBJECT_CLASS_ID", "0"))
object_tracker = ByteTracker(
track_high_thresh=TRACK_HIGH_THRESH,
track_low_thresh=TRACK_LOW_THRESH,
match_thresh=TRACK_MATCH_THRESH,
track_buffer=TRACK_BUFFER,
min_hits=TRACK_MIN_HITS,
)
# Per-track zone state: enter left → LEFT feeder; enter right → RIGHT feeder.
# Keys survive ID switches via inheritance (see _inherit_prev).
object_zone_state = {}
recent_cross_left = deque()
recent_cross_right = deque()
# Monotonic timestamp of last successful count per feeder (for cooldown).
last_count_mono = {"left": None, "right": None}
detect_snapshot_ids = set()
object_cross_flash = {}
zone_pulse = count_left_pulse = count_right_pulse = 0
popups = []
session_start = time.time()
frame_idx = 0
inf_ms = 0.0
prev_gray = None
frames_since_infer = 0
video_writer = None
crossing_times = deque()
counter_left = 0
counter_right = 0
last_snapshot_cleanup = 0.0
counting_active = True
last_control_poll = 0.0
control_socket = None
if CONTROL_ENABLED:
if not Path(CONTROL_FILE).exists():
write_control_file(CONTROL_DEFAULT_COUNTING)
counting_active = read_counting_flag(CONTROL_DEFAULT_COUNTING)
print(
f"Runtime control enabled | file={CONTROL_FILE} | "
f"counting={'ON' if counting_active else 'OFF'}"
)
if CONTROL_SOCKET_ENABLED:
try:
control_socket = start_control_socket()
except Exception as exc:
print(f"[{now_str()}] Failed to start control socket: {exc}")
cap, w, h, fps = connect_stream(SOURCE)
if cap is None:
store.shutdown()
model.release()
return
zone_left = resolve_zone_left(w, h)
zone_right = resolve_zone_right(w, h)
print(
f"RKNN+ByteTrack zone counter | {w}x{h} @ {fps}fps | "
f"LEFT feeder {zone_left} RIGHT feeder {zone_right}"
)
print(
f"Cooldownoldown: left={ZONE_COOLDOWN_LEFT_SEC}s right={ZONE_COOLDOWN_RIGHT_SEC}s "
f"(0=off) | Dwell: left={ZONE_DWELL_LEFT_SEC}s right={ZONE_DWELL_RIGHT_SEC}s "
f"(0=immediate)"
)
print(f"Model: {MODEL_PATH} | imgsz={IMGSZ} | core_mask={CORE_MASK}")
print(
f"ByteTrack: high_thresh={TRACK_HIGH_THRESH} low_thresh={TRACK_LOW_THRESH} "
f"match_thresh={TRACK_MATCH_THRESH} buffer={TRACK_BUFFER}"
)
print(f"DB: {DB_PATH}")
print(f"State: {STATE_FILE}")
if RECORD_VIDEO:
video_writer = VideoSegmentWriter(OUTPUT_DIR, w, h, fps, VIDEO_SEGMENT_SEC)
reconnect_count = 0
while not shutdown_requested:
ret, frame = cap.read()
if not ret:
if not IS_LIVE:
break
reconnect_count += 1
print(
f"Stream dropped (attempt {reconnect_count}), reconnecting in {RECONNECT_DELAY_SEC}s..."
)
cap.release()
time.sleep(RECONNECT_DELAY_SEC)
cap, w, h, fps = connect_stream(SOURCE)
if cap is None:
break
zone_left = resolve_zone_left(w, h)
zone_right = resolve_zone_right(w, h)
continue
now = time.time()
elapsed = now - session_start
mono = time.monotonic()
object_crossed_frame = False
cross_events_frame = []
detect_events_frame = []
if CONTROL_ENABLED and (now - last_control_poll) >= CONTROL_POLL_SEC:
last_control_poll = now
new_flag = read_counting_flag(CONTROL_DEFAULT_COUNTING)
if new_flag != counting_active:
counting_active = new_flag
print(f"[{now_str()}] Counting {'RESUMED' if counting_active else 'PAUSED'} via control file")
skip_inference = False
if not counting_active:
skip_inference = True
elif MOTION_DETECTION_ENABLED:
gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
if prev_gray is not None:
diff = cv2.absdiff(gray, prev_gray)
moved = int(np.count_nonzero(diff > MOTION_PIXEL_DELTA))
moved_frac = moved / diff.size
skip_inference = moved_frac < MOTION_MIN_AREA_FRAC
if frames_since_infer >= MOTION_HEARTBEAT_FRAMES:
skip_inference = False
prev_gray = gray
detections = []
if not skip_inference:
frames_since_infer = 0
inf_start = time.time()
detections = model(frame)
inf_ms = inf_ms * 0.9 + (time.time() - inf_start) * 1000 * 0.1
else:
frames_since_infer += 1
object_boxes_xyxy = []
object_scores = []
object_kpts_list = []
object_cx_list = []
object_cy_list = []
for det in detections:
bbox = det["bbox"]
score = det["score"]
cls_id = det["cls"]
kpts = det["keypoints"]
cx = (bbox[0] + bbox[2]) / 2.0
cy = (bbox[1] + bbox[3]) / 2.0
if cls_id == object_cls:
object_boxes_xyxy.append(bbox)
object_scores.append(score)
object_kpts_list.append(kpts)
object_cx_list.append(cx)
object_cy_list.append(cy)
object_boxes_xyxy = np.array(object_boxes_xyxy, dtype=np.float32).reshape(-1, 4)
object_scores = np.array(object_scores, dtype=np.float32)
object_track_map, object_det_to_track, object_lost_map = object_tracker.update(
object_boxes_xyxy, object_scores
)
if DEBUG_TRACKING and len(object_boxes_xyxy) > 0:
scores_str = (
f" scores: {object_scores.round(3).tolist()}"
if len(object_scores) > 0
else ""
)
tracks_str = (
f" det->track: {dict(object_det_to_track)}"
if object_det_to_track
else ""
)
zone_str = ""
if object_zone_state:
left_ids = sorted(
tid
for tid, st in object_zone_state.items()
if st.get("side") == "left"
)
right_ids = sorted(
tid
for tid, st in object_zone_state.items()
if st.get("side") == "right"
)
if left_ids:
zone_str += f" left_counted: {left_ids}"
if right_ids:
zone_str += f" right_counted: {right_ids}"
print(
f"[DEBUG F{frame_idx}] dets={len(object_boxes_xyxy)} "
f"tracks={len(object_track_map)} "
f"zone_L={zone_left} zone_R={zone_right}{scores_str}"
f"{tracks_str}{zone_str}",
flush=True,
)
# --- zone dwell counting only on DETECTED objects this frame ---
# A sack must remain inside a feeder zone for ZONE_DWELL_*_SEC before
# it is counted (and a cross snapshot is saved). Leaving the zone
# resets the dwell timer. Cooldown still applies after a successful count.
for di in range(len(object_boxes_xyxy)):
tid = object_det_to_track.get(di)
if tid is None:
continue
cx = object_cx_list[di]
cy = object_cy_list[di]
if tid not in detect_snapshot_ids:
detect_snapshot_ids.add(tid)
detect_events_frame.append(tid)
if tid not in object_tracked:
inherited = _inherit_prev(
object_tracked, tid, cx, cy, mono, INHERIT_SEC, INHERIT_PX
)
if inherited is not None:
object_tracked[tid] = inherited[:3]
src_tid = inherited[3]
if src_tid in object_zone_state:
object_zone_state[tid] = _copy_zone_state(
object_zone_state[src_tid]
)
if DEBUG_TRACKING:
src_st = object_zone_state.get(tid, {})
print(
f"[DEBUG F{frame_idx}] INHERIT prev for new tid={tid} "
f"from tid={src_tid} ({inherited[0]:.1f},{inherited[1]:.1f})"
f" counted={src_st.get('counted', False)}"
f" side={src_st.get('side')}"
f" dwell={src_st.get('dwell_feeder')}",
flush=True,
)
if tid not in object_zone_state:
object_zone_state[tid] = _default_zone_state()
st = object_zone_state[tid]
feeder = resolve_feeder_at(cx, cy, zone_left, zone_right)
if st["counted"]:
object_tracked[tid] = (cx, cy, mono)
continue
if feeder is None:
# Left all zones — reset dwell progress.
if st["dwell_feeder"] is not None and DEBUG_TRACKING:
print(
f"[DEBUG F{frame_idx}] DWELL RESET: tid={tid} "
f"left feeder={st['dwell_feeder']}",
flush=True,
)
st["dwell_feeder"] = None
st["dwell_since"] = None
object_tracked[tid] = (cx, cy, mono)
continue
if st["dwell_feeder"] != feeder:
st["dwell_feeder"] = feeder
st["dwell_since"] = mono
if DEBUG_TRACKING:
print(
f"[DEBUG F{frame_idx}] DWELL START: tid={tid} "
f"feeder={feeder} need={feeder_dwell_sec(feeder):.1f}s",
flush=True,
)
object_tracked[tid] = (cx, cy, mono)
continue
dwell_need = feeder_dwell_sec(feeder)
dwell_elapsed = mono - (st["dwell_since"] or mono)
if dwell_need > 0 and dwell_elapsed < dwell_need:
object_tracked[tid] = (cx, cy, mono)
continue
# Dwell complete — apply dedup / cooldown, then count + snapshot.
recent = recent_cross_left if feeder == "left" else recent_cross_right
if is_duplicate_cross(recent, cx, cy, frame_idx):
st["counted"] = True
st["side"] = feeder
if DEBUG_TRACKING:
print(
f"[DEBUG F{frame_idx}] DUP ZONE IGNORED: tid={tid} "
f"cx={cx:.1f} cy={cy:.1f} feeder={feeder}",
flush=True,
)
object_tracked[tid] = (cx, cy, mono)
continue
if feeder_in_cooldown(feeder, mono, last_count_mono):
# Keep dwelling; count once cooldown expires if still in zone.
if DEBUG_TRACKING:
cool = feeder_cooldown_sec(feeder)
elapsed = mono - (last_count_mono.get(feeder) or mono)
print(
f"[DEBUG F{frame_idx}] DWELL WAIT COOLDOWN: tid={tid} "
f"feeder={feeder} elapsed={elapsed:.2f}s / {cool}s",
flush=True,
)
object_tracked[tid] = (cx, cy, mono)
continue
if DEBUG_TRACKING:
print(
f"[DEBUG F{frame_idx}] DWELL COUNT: tid={tid} "
f"cx={cx:.1f} cy={cy:.1f} feeder={feeder} "
f"dwell={dwell_elapsed:.2f}s",
flush=True,
)
recent.append((frame_idx, cx, cy))
st["counted"] = True
st["side"] = feeder
last_count_mono[feeder] = mono
if feeder == "left":
counter_left += 1
count_left_pulse = COUNT_PULSE_FRAMES
feeder_total = counter_left
else:
counter_right += 1
count_right_pulse = COUNT_PULSE_FRAMES
feeder_total = counter_right
# Frigate-style count log, e.g.
# "Total karung di kandang_bawah_feeder_kanan: 5"
print(
f"Total {OBJECT_LABEL} di {feeder_display_name(feeder)}: {feeder_total}",
flush=True,
)
store.record_zone_entry(tid, feeder)
if cross_logger:
cross_logger.write_row(
[
store.get_counting_date(),
frame_idx,
feeder,
tid,
]
)
object_crossed_frame = True
cross_events_frame.append((tid, feeder))
crossing_times.append(mono)
object_cross_flash[tid] = CROSS_FLASH_FRAMES
popups.append(
{
"x": int(cx) - 12,
"y": int(cy),
"born": frame_idx,
"text": f"+1 {feeder.upper()}",
}
)
object_tracked[tid] = (cx, cy, mono)
for tid, (cx, cy) in object_lost_map.items():
if tid not in object_tracked:
object_tracked[tid] = (cx, cy, mono)
for di in range(len(object_boxes_xyxy)):
tid = object_det_to_track.get(di)
if tid is None:
continue
bbox = object_boxes_xyxy[di]
x1, y1, x2, y2 = int(bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3])
flash = object_cross_flash.get(tid, 0)
color = C_GREEN if flash > 0 else C_OBJECT_BOX
cv2.rectangle(frame, (x1, y1), (x2, y2), color, 3 if flash > 0 else 2)
draw_pill(frame, f"ID {tid}", x1, y1 - 4, color)
kpts = object_kpts_list[di] if di < len(object_kpts_list) else None
if kpts is not None:
draw_skeleton_bold(frame, kpts)
if object_crossed_frame:
zone_pulse = ZONE_PULSE_FRAMES
display_total = store.display_total()
while crossing_times and mono - crossing_times[0] > RATE_WINDOW_SEC:
crossing_times.popleft()
rate = (len(crossing_times) / RATE_WINDOW_SEC * 60) if crossing_times else 0.0
draw_elegant_zone(
frame, zone_left, zone_pulse, label="LEFT FEEDER", color=C_ZONE_LEFT
)
draw_elegant_zone(
frame, zone_right, zone_pulse, label="RIGHT FEEDER", color=C_ZONE_RIGHT
)
draw_zone_count(
frame, zone_left, counter_left, "LEFT", C_ZONE_LEFT, count_left_pulse
)
draw_zone_count(
frame, zone_right, counter_right, "RIGHT", C_ZONE_RIGHT, count_right_pulse
)
draw_hud(
frame,
w,
counter_left + counter_right,
counter_left,
counter_right,
elapsed,
rate,
)
draw_footer(
frame,
w,
h,
frame_idx,
"LIVE" if IS_LIVE else "FILE",
inf_ms,
Path(MODEL_PATH).name,
)
popups = draw_popups(frame, popups, frame_idx)
if CONTROL_ENABLED and not counting_active:
badge = "COUNTING PAUSED"
(bw, bh), _ = cv2.getTextSize(badge, cv2.FONT_HERSHEY_SIMPLEX, 0.6, 2)
bx = w // 2 - bw // 2
overlay_rect(frame, bx - 14, 48, bx + bw + 14, 48 + bh + 18, C_PANEL, alpha=0.75)
cv2.rectangle(frame, (bx - 14, 48), (bx + bw + 14, 48 + bh + 18), C_ACCENT, 2)
cv2.putText(frame, badge, (bx, 48 + bh + 6), cv2.FONT_HERSHEY_SIMPLEX, 0.6, C_ACCENT, 2, cv2.LINE_AA)
for tid in list(object_cross_flash):
object_cross_flash[tid] -= 1
if object_cross_flash[tid] <= 0:
del object_cross_flash[tid]
zone_pulse = max(0, zone_pulse - 1)
count_left_pulse = max(0, count_left_pulse - 1)
count_right_pulse = max(0, count_right_pulse - 1)
if video_writer is not None:
video_writer.write(frame)
if LIVE_STREAM_ENABLED and frame_idx % LIVE_STREAM_EVERY_N == 0:
try:
Path(LIVE_STREAM_FRAME_PATH).parent.mkdir(parents=True, exist_ok=True)
ok, jpeg = cv2.imencode(
".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, LIVE_STREAM_QUALITY]
)
if ok:
tmp_path = f"{LIVE_STREAM_FRAME_PATH}.tmp"
with open(tmp_path, "wb") as f:
f.write(jpeg.tobytes())
os.replace(tmp_path, LIVE_STREAM_FRAME_PATH)
except Exception:
pass
if (SAVE_DETECT_SNAPSHOT and detect_events_frame) or (
SAVE_CROSS_SNAPSHOT and cross_events_frame
):
try:
ts = datetime.now().strftime("%Y%m%d_%H%M%S_%f")[:-3]
if SAVE_DETECT_SNAPSHOT and detect_events_frame:
detect_dir = Path(CROSS_SNAPSHOT_DIR) / "detect"
detect_dir.mkdir(parents=True, exist_ok=True)
for tid in detect_events_frame:
fname = f"{ts}_detect_id{tid}_f{frame_idx}.jpg"
cv2.imwrite(
str(detect_dir / fname),
frame,
[cv2.IMWRITE_JPEG_QUALITY, CROSS_SNAPSHOT_QUALITY],
)
if SAVE_CROSS_SNAPSHOT and cross_events_frame:
cross_dir = Path(CROSS_SNAPSHOT_DIR) / "cross"
cross_dir.mkdir(parents=True, exist_ok=True)
for tid, feeder in cross_events_frame:
fname = f"{ts}_{feeder}_id{tid}_f{frame_idx}.jpg"
cv2.imwrite(
str(cross_dir / fname),
frame,
[cv2.IMWRITE_JPEG_QUALITY, CROSS_SNAPSHOT_QUALITY],
)
if now - last_snapshot_cleanup >= CROSS_SNAPSHOT_CLEANUP_SEC:
cleanup_snapshots(
CROSS_SNAPSHOT_DIR,
CROSS_SNAPSHOT_MAX_FILES,
CROSS_SNAPSHOT_MAX_AGE_DAYS,
)
last_snapshot_cleanup = now
except Exception as exc:
print(f"[{now_str()}] Failed to save snapshot: {exc}")
frame_idx += 1
prune_stale_tracks(object_tracked, mono)
cap.release()
if video_writer is not None:
video_writer.release()
if cross_logger:
cross_logger.close()
if control_socket is not None:
try:
control_socket.close()
except Exception:
pass
model.release()
store.shutdown()
print("\n=== Daily Counter Summary (SQLite) ===")
print(f"Database: {DB_PATH}")
object_tracked = {}
if __name__ == "__main__":
run()