Introdction

This is a multiple object tracking project that use Yolov3(opencv ver.) as detector and DeepSort as trackers.

Different from original method in Deepsort. I did some improvement.

1. Remove the network and reuse the feature maps from Yolov3 with following methods:

   1.1 : global average pooling

   1.2 : global max pooling

   1.3 : Part-based local average pooling

   1.4 : Part-based local max pooling

2. Extend the method from only tracking poeple to all the objects that the detector detected.

Videos demo on Cityscapes (record at 1 fps)

1. Video 1
2. Video 2
3. Video 3

Installation

1. Clone the repository:

git clone https://github.com/world4jason/tracking_by_detection_with_yolo_deep_sort

2. Download pre-generated detections and the CNN checkpoint file from here and the netowrk folder under model_data

3. Download YOLOv3 weights from Yolo Website and put into model_data/yolov3 folder or runs

sh tools/get_yolo.sh

Dependencies

See requirements.txt

Additionally, network feature generation for original Deepsort method requires TensorFlow-1.4.0. or TensorFlow-1.5.0.

Running

python3 demo.py

Parameters

usage: demo.py [-h] [--conf_threshold CONF_THRESHOLD]
               [--nms_threshold NMS_THRESHOLD] [--net_width NET_WIDTH]
               [--net_height NET_HEIGHT] [--tracker_type TRACKER_TYPE]
               [--split SPLIT] [--tracker_nms_threshold TRACKER_NMS_THRESHOLD]
               [--max_cosine_distance MAX_COSINE_DISTANCE]
               [--nn_budget NN_BUDGET] [--label_path LABEL_PATH]
               [--model_path MODEL_PATH] [--weight_path WEIGHT_PATH]
               [--sort_model_path SORT_MODEL_PATH]
               [--output_video OUTPUT_VIDEO] [--video_path VIDEO_PATH]

YoloV3 with Variants Sorts

optional arguments:
  -h, --help            show this help message and exit
  --conf_threshold CONF_THRESHOLD
  --nms_threshold NMS_THRESHOLD
  --net_width NET_WIDTH
  --net_height NET_HEIGHT
  --tracker_type TRACKER_TYPE
  --split SPLIT
  --tracker_nms_threshold TRACKER_NMS_THRESHOLD
  --max_cosine_distance MAX_COSINE_DISTANCE
  --nn_budget NN_BUDGET
  --label_path LABEL_PATH
  --model_path MODEL_PATH
  --weight_path WEIGHT_PATH
  --sort_model_path SORT_MODEL_PATH
  --output_video OUTPUT_VIDEO
  --video_path VIDEO_PATH

License

world4jason/tracking_by_detection is released under the GPL-3.0.