End to End Detection and Tracking

The repository contains code for detection and tracking. The code uses Deep Learning Detectors and Kalman Filter for tracking.

Sample Video

Usage

python main.py --detector yolov3 --input <input_video> --output <output_path>

Setup

git clone 

# to use yolo-v3 detector
git submodule update --init --recursive

# download yolo-v3 weights
cd detection/yolov3
mkdir weights
cd weights
wget https://pjreddie.com/media/files/yolov3.weights 

Detectors

Currently, module supports two detectors:

• Mobilenet Single Shot Detector
• Yolo-v3

One can implement their own detector by extending BaseDetector class defined here

Trackers

Currently, module supports only Kalman Filter based tracker: KalmanTracker.

One can implement their own tracker by extending BaseTracker class defined here

Implementation Methodology

The class DetectAndTrack, defined here, maintains list of currently tracked objects.

• Process current frame to obtain new detections
• Assign current detections to existing trackers using Hungarian Algorithm. This would result in matches, unmatched detections and unmatched trackers
• Assign new trackers to unmatched detections
• Keep old trackers for consecutive unmatched detections for max_age frames
• Update tracker's state using tracking algorithm (currently Kalman Filter)