A cost-effective, reliable and accurate framework for multiple-target tracking by detection approach using deep neural network

Abstract

Over the years the area of object tracking and detection has emerged and become ubiquitous owing to its potential contribution towards video surveillance applications. Multiple object tracking (MOT) estimates the trajectory of several objects of interest simultaneously over time in a series of video frames. Even though various research proposals have encouraged the use of machine learning techniques in designing multi-object trackers, the existing solutions need to be more practicable for online tracking due to more complicated algorithms, The study, therefore, introduces a cost-effective tracking solution for multiple–target tracking by detection where it incorporates the you only look once version 4 (YOLOv4) and person re-identification network, which are further integrated with the proposed tracking model, which considers both bounding box and appearance features to handle the motion prediction and data association respectively. The novelty of this approach lies in considering appearance features, which not only help predict tracks through allocations problem solving but also handle the cost of computation problems. Here, the system utilizes a pre-trained association metric with which the occlusion challenges are also handled, whereas the target tracking has taken place even in more extended periods of occlusion, making it suitable with the existing efficient tracking algorithms.