Framework for Pedestrian Detection, Tracking and Re-identification in Video Surveillance System

In this work, we present a framework for implementing pedestrian re-identification in real world surveillance camera systems. Pedestrian re-identification has seen a lot of attention as a stand-alone identifier that takes in cropped images of pedestrians and matches their IDs. The focus of this work is to produce the input needed for the re-identifier network form a raw video output from the surveillance cameras. The integrated system utilizes Tensorflow object detection API with a Faster-RCNN pre-trained model, for pedestrian detection, a discriminative correlation filter from [1], for the purpose of short-term tracking of subjects in each camera and a pedestrian re-identification network from [2]. This configuration resulted in an accuracy above 93% on our dataset, that was automatically produced using pedestrian detection and tracking algorithms.