tieubaoca/CiviWatch-Traffic-Models

CiviWatch Traffic Models (Vietnam Urban CCTV)

Suite of 3 fine-tuned traffic & urban safety object detectors built for CiviWatch v2 on 25,977 HD frames (299,157 boxes) from Ho Chi Minh City urban CCTV cameras (AIC_HCMC_2020) mixed with balanced pedestrian and obstacle annotations.

Repository code: remiehneppo/ai-camera (branch: civi-watch)

---

1. Included Models & Licenses

Subdirectory	Model	Input Size	Decode Mode	File Sizes	License	Recommended Hardware
yolox-tiny-416/	CiviWatch YOLOX-Tiny	416x416	yolox_grid	ONNX 20.2 MB / .pth 39 MB	Apache-2.0	CPU-only, Raspberry Pi 5, Rockchip RK3566/RK3588 NPU
yolox-s-640/	CiviWatch YOLOX-s	640x640	yolox_grid	ONNX 35.8 MB / .pth 69 MB	Apache-2.0	Commercial Edge AI Box / GPU needing permissive Apache-2.0
yolo26s-640/	CiviWatch YOLO26s	640x640	nms_free ([1, 300, 6])	ONNX 38.2 MB / .pt 20.3 MB	AGPL-3.0	Jetson Orin NX / GPU Server / Fast NMS-free CPU

---

2. Class Schema (6 Classes)

{
  "0": "person",
  "1": "two_wheeler",
  "2": "car",
  "3": "bus",
  "4": "truck",
  "5": "obstacle_object"
}

---

3. Benchmark Results (AIC_HCMC_2020 Validation & Full-Edge Video Suite)

Evaluated directly inside CiviWatchDetector (ONNXRuntime CPUExecutionProvider) on 100 HCMC validation frames across day/night intersections:

Model	two_wheeler Recall@0.4	car Recall@0.4	bus Recall@0.4	truck Recall@0.4	Overall Recall@0.4	Night Recall	Validation mAP50 (3,858 imgs)	CPU P50 Latency
YOLOX-Tiny (COCO Baseline)	14.2%	54.2%	31.2%	35.7%	22.7%	4.8%	—	21.1 ms
CiviWatch YOLOX-Tiny (416)	90.5%	84.9%	81.2%	94.3%	89.5%	91.9%	82.1%	20.6 ms
CiviWatch YOLOX-s (640)	94.4%	88.6%	87.5%	97.1%	93.4%	93.5%	88.6%	93.7 ms
CiviWatch YOLO26s (640)	95.6%	91.6%	93.8%	97.1%	95.0%	91.9%	90.2%	46.5 ms

Full-Edge Video Pipeline Throughput (LivePerceptionRunner + ByteTrack + Homography + Outbox)

Model	sample_01.mp4 (HCMC 720p)	sample_02.mp4 (HCMC 720p)	shibuya_crossing_1080p.mp4
CiviWatch YOLOX-Tiny (416)	24.3 FPS (1,642 dets)	21.5 FPS (1,973 dets)	19.6 FPS (1,640 dets)
CiviWatch YOLOX-s (640)	12.0 FPS (1,828 dets)	10.8 FPS (1,926 dets)	10.0 FPS (1,635 dets)
CiviWatch YOLO26s (640 NMS-Free)	14.8 FPS (1,648 dets)	12.8 FPS (1,911 dets)	10.3 FPS (1,235 dets)

---

4. Quick Usage with CiviWatch

Download a model folder and pass its manifest.json and .onnx path to run_edge.py or CiviWatchDetector:

huggingface-cli download tieubaoca/CiviWatch-Traffic-Models --local-dir models/hf-civiwatch

# Run Edge Pipeline with YOLOX-Tiny HCMC (Apache-2.0, ~24 FPS CPU):
python run_edge.py --module-config fixtures/sample_module_config.json \
    --source file --file sample_01.mp4 \
    --manifest models/hf-civiwatch/yolox-tiny-416/manifest.json \
    --model models/hf-civiwatch/yolox-tiny-416/yolox_tiny_civiwatch_416.onnx

# Run Edge Pipeline with YOLO26s HCMC (NMS-Free, 95.6% motorcycle recall):
python run_edge.py --module-config fixtures/sample_module_config.json \
    --source file --file sample_01.mp4 \
    --manifest models/hf-civiwatch/yolo26s-640/manifest.json \
    --model models/hf-civiwatch/yolo26s-640/yolo26s_civiwatch_640.onnx

Preprocessing Notes

• YOLOX-Tiny / YOLOX-s (yolox_grid): Top-left letterbox (pad=114), BGR, float32 range [0.0, 255.0] (scale=1.0).
• YOLO26s (nms_free): Top-left letterbox (pad=114), RGB, float32 range [0.0, 1.0] (scale=1/255.0), output tensor shape [1, 300, 6] as [x1, y1, x2, y2, score, class_id].