ktzsh/object-tracking

Multiple Object Tracking System in Keras + (Detection Network - YOLO)

Object Tracking

1. Simultaneous Multiple Object Detection and Tracking System in Keras (Detection network based on YOLOv2 - reimplemented in keras)
2. Single Object Tracking with FasterRCNN and YOLOv2/v3 as detection backends

Trackers Available:

1. TinyTracker (Static Detection Priors from FasterRCNN or Yolov2/v3)[Single Object]:
2. TinyHeatmapTracker (Static Detection Priors from FasterRCNN or Yolov2/v3)[Single Object]:
3. MultiObjDetTracker (Trainable Detection Priors from Yolov2 reimplmented in Keras)[Multiple Objects]:

NOTE: Yolov2 reimplementation in Keras as standalone detector also available

Dependencies

1. Tensorflow
2. Keras
3. OpenCV
4. easydict (for py-faster-rcnn)
5. cython (for py-faster-rcnn)
6. imgaug

Instllation

1. Run git clone --recursive https://github.com/kshitiz38/object-tracking.git

   • NOTE: If you didn't clone with the --recursive flag run manually the following code git submodule update --init --recursive

2. Darknet

   • Follow instructions at https://pjreddie.com/darknet/install/ cd darknet && make
   • NOTE: I recommend disabling CUDNN in Makefile since it gives strange results depending on your version

3. Faster RCNN

   • Follow intructions at https://github.com/rbgirshick/py-faster-rcnn#installation-sufficient-for-the-demo

     cd py-faster-rcnn
     cd lib && make && cd ../
     cd caffe-fast-rcnn && mkdir build && cd build && cmake ..
     make all && make install
     

Usage

1. For Single Object Tracking
   1. Modify Parameters in config.jon
   2. Convert Datasets to PASCAL VOC format if not already
      • Run python utility/tb_to_pascal.py' or 'python utility/tb_to_pascal.py or write one for your own dataset
   3. Run python trainer.py
2. For Simultaneous Multiple Object Detection and Tracking
   1. Modify Parameters in KerasYOLO.py and MultiObjDetTracker.py
   2. Convert Datasets like above specify paths in MultiObjDetTracker.py already done for ImageNet Vid and MOT17
   3. Run python trainer.py

NOTE :

• Call single_object_tracking() in trainer.py for Single Object Detection with fixed detection priors from Other Detection backends
• Call simult_multi_obj_detection_tracking() in trainer.py for Simultaneous Multiple Object Detction and Tracking with Yolov2 Reimplemented in Keras

Model Architectures

• Coming Soon!!
• Feel free to figure out yourself!! See models_tracking and models_detection directories

TODOs

• Add theory and model architectures explaination
• Add config.json file for parameters for MultiObjDetTracker and KerasYOLO
• Benchmark for ImagenetVid Challenge, MOT and VisualTB Datasets
• Add support for Detectron models as detection backend

References

1. https://github.com/Guanghan/ROLO
2. https://github.com/experiencor/keras-yolo2