Computer Vision - Object Tracking

Sources

Algorithm for Object Tracking

Deep SORT Implementation Overview

YOLOv3 Network

Generating the bounding boxes for detected objects

yolo3image

Detections

Encode given cropped image

Kalman Filter

Used to generate predictions of objects' future locations

kal

probability

IoU Matching

Solve linear assignment problem as well as cascade matching

Nearest Neighbor Matching

Solves cosine distance and Eucladian distance for Deep Appearance Descriptor as well as Kalman filter

Tracking

Combines modules to create final object

Object Tracking

Complete and run algorithm in a loop for each video.