Vehicle Counting/Tracking using EfficientDet + DeepSORT (Pytorch)

Source code of #Team 059 for AI City Challenge of Ho Chi Minh City 2020
- Task info: http://aichallenge.hochiminhcity.gov.vn/huong-dan-nhom-1
- Leaderboard: http://aichcmc.ml/

Environments

pytorch 1.6.0
Windows 10

Demo

File Structure:

this repo
│   detect.py
│   track.py
│   train.py
│
└───configs
│      aic-hcmc.yaml
│      cam_configs.yaml
│
└───datasets  
│   │
│   └───aic-hcmc2020
│       │
│       └───images
│       │     00000.jpg
│       │     00001.jpg
│       │     ....
│       └───annotations
│       │     instances_train.json
│       │     instances_val.json
│       │
│    └───annotations
│    |      cam_01.json
│    └───videos
│        │   
│        │  cam_01.mp4
│        │  cam_02.mp4
│        │  ...
│

Dataset:

AIC-HCMC-2020: not public

Pretrained weights:

Download pretrained EfficientDet from original repo
Download finetuned weights on AIC-HCMC-2020 dataset:

Model	Image Size	Weights	MAP-S	MAP-M	MAP-L	ALL
EfficientDet-D2	768 x 768	link	0.068	0.312	0.390	0.32

Custom dataset:

To use on custom dataset, see example yaml file in ./configs directory and create a new one and specified:

  project_name: <name of project>  # also the folder name of the dataset that under datasets folder
  train_imgs: <training images directory>
  val_imgs: <val images directory>
  train_anns: <training images json annotations>
  val_anns: <val images json annotations>
  obj_list: <list of string contains object name>
  ...

Method:

Training

Split the video into frames using preprocess.py
To manually label, use labelme
Download all pretrained weights from above
Use EfficientDet for vehicle detection task, finetune the model on the labeled dataset using train.py
Use DeepSORT for car tracking, not need to retrain this model, only inference

Inference

Use trained detection model to detect vehicle in the video using detect.py, all the bounding boxes, classes prediction will be saved to json files
Use pretrained tracking model to track vehicle by their results from detection model above, use track.py

Finetuning EfficientDet:

python train.py -c=<version number of EfficientDet> --config=<path to project config yaml file>

Extra Parameters:
- --resume: path to checkpoint to resume training
- --batch_size: batch size, recommend 4 - 8
- --head_only: if train only the head
- --num_epochs: number of epochs
- --saved_path: path to save weight
- --val_interval: validate per number of epochs
- --save_interval: save per number of iterations
- --log_path: tensorboard logging path

Inference on AIC-HCMC testset:

Run detection for detecting bounding boxes and classes confidence scores

python detect.py -c=<version of EfficientDet> --config=<path to project config yaml file> --video_path=<path to video .mp4>

Extra Parameters:
- -c: version of EfficientDet
- --config: configs for detection
- --batch_size: batch size, recommend 4 - 8
- --min_conf: minimum confident for detection
- --min_iou: minimum iou for detection
- --weight: pretrained weight
- --saved_path: path to save detection results
- --output_path: path to save mp4 file

Run tracking on detected bounding boxes and classes confidence scores

python track.py video_name --config=<cam configs> --out_path=<path to save results>

Parameters:
- video_name: video name that is defined in config
- --config: config for all cam
- --output_path: path for output video and counting results
- --display: for output video