SJTU EI339 Artificial Intelligence course project: Multi-Object Tracking based on Deep SORT.
For example, the tracking result of MOT16-02
is: (click to watch it on Youtube)
Anaconda (or Miniconda) is highly recommended for this project.
- Install CenterNet according to its instructions
- Install necessary packages for deep SORT with
pip install -r requirements.txt
- Get pretrained models:
- CenterNet's models can be found in the model zoo, and they are supposed to be put under
CenterNet/models
- Deep SORT's models can be downloaded from Google Drive and then put under
deep_sort/deep/checkpoint
- CenterNet's models can be found in the model zoo, and they are supposed to be put under
mot_challenge.py
takes the path to a MOT Challenge sequence and produces the tracking result in a text file (its format is consistent with MOT Challenge requirements for evaluation).
python mot_challenge.py [-h]
[--model_path MODEL_PATH]
[--arch ARCH]
[--deepsort_checkpoint DEEPSORT_CHECKPOINT]
[--output_file OUTPUT_FILE]
[--min_confidence MIN_CONFIDENCE]
[--min_detection_height MIN_DETECTION_HEIGHT]
[--nms_max_overlap NMS_MAX_OVERLAP]
[--max_cosine_distance MAX_COSINE_DISTANCE]
[--debug]
[--debug_dir DEBUG_DIR]
[--no_cuda]
sequence_dir
demo_video.py
takes the path to a video and produces the tracking result in a video for visualization.
python demo_video.py [-h]
[--model_path MODEL_PATH]
[--arch ARCH]
[--deepsort_checkpoint DEEPSORT_CHECKPOINT]
[--output OUTPUT]
[--min_confidence MIN_CONFIDENCE]
[--max_cosine_distance MAX_COSINE_DISTANCE]
[--no_cuda]
video_path
- SORT: Simple Online and Realtime Tracking, paper, code
- Deep SORT: Simple Online and Realtime Tracking with a Deep Association Metric, paper, code
- CenterNet: Objects as Points, paper, code
- Deep SORT with YOLOv3: ZQPei/deep_sort_pytorch
- MOTChallenge: The Multiple Object Tracking Benchmark, MOTChallenge