🏍️ 🚙 Vehicle Tracking using YOLOv5 + DeepSORT 🚌

Dev logs

[19/12/2021] Update to new YOLOv5 version 6. Can load checkpoints from original repo now 🤞
[16/07/2021] BIG REFACTOR Code is cleaned and working fine now, promise 🤞
[27/09/2021] All trained checkpoints on AIC-HCMC-2020 have been lost. Now use pretrained models on COCO for inference.

Method

• Use YOLOv5 for vehicle detection task, only considers objects in Region of Interest (ROI)
• Use DeepSORT for car tracking, not need to retrain this model, only inference
• Use Cosine Similarity to assign object's tracks to most similar directions.
• Count each type of vehicle on each direction.

📔 Notebook

• For inference, use this notebook
• To retrain detection model, follow instructions from original Yolov5

---

Dataset

• AIC-HCMC-2020: link
• Direction and ROI annotation format:

cam_01.json # match video name
{
    "shapes": [
        {
            "label": "zone",
            "points": [[x1,y1], [x2,y2], [x3,y3], [x4,y4], ... ] #Points of a polygon
        },
        {
            "label": "direction01",
            "points": [[x1,y1], [x2,y2]] #Points of vector
        },
        {
            "label": "direction{id}",
            "points": [[x1,y1], [x2,y2]]
        },...
    ],
}

🥇 Pretrained weights

• Download finetuned models from on AIC-HCMC-2020 dataset:

Model	Image Size	Weights	Precision	Recall	[email protected]	[email protected]
YOLOv5s	640x640	link	0.87203	0.87356	0.91797	0.60795
YOLOv5m	1024x1024	link	0.89626	0.91098	0.94711	0.66816

🌟 Inference

• File structure

this repo
│   detect.py
└───configs
│      configs.yaml           # Contains model's configurations
│      cam_configs.yaml       # Contains DEEPSORT's configuration for each video

• Install dependencies by pip install -r requirements.txt
• To run full pipeline:

python run.py --input_path=<input video or dir> --output_path=<output dir> --weight=<trained weight>

• Extra Parameters:
  • --min_conf: minimum confident for detection
  • --min_iou: minimum iou for detection

---

Results

• After running, a .csv file contains results has following example format:

track_id	frame_id	box	color	label	direction	fpoint	lpoint	fframe	lframe
2	3	[607, 487, 664, 582]	(144, 238, 144)	0	1	(635.5, 534.5)	(977.0, 281.5)	3	109
2	4	[625, 475, 681, 566]	(144, 238, 144)	0	1	(635.5, 534.5)	(977.0, 281.5)	3	109
2	5	[631, 471, 686, 561]	(144, 238, 144)	0	1	(635.5, 534.5)	(977.0, 281.5)	3	109

• With:
  • track_id: the id of the object
  • frame_id: the current frame
  • box: the box wraps around the object in the corresponding frame
  • color: the color which is used to visualize the object
  • direction: the direction of the object
  • fpoint, lpoint: first/last coordinate where the object appears
  • fframe, lframe: first/last frame where the object appears

Visualization result

References

• DeepSORT from https://github.com/ZQPei/deep_sort_pytorch
• YOLOv5 from https://github.com/ultralytics/yolov5