A Real-time and Online Multiple-Type Object Tracking Method with Deep Features

Abstract

Object tracking is one of the most important things in intelligent vision system. Meanwhile, the most challenging issue in object tracking is how to keep the target's identity unchangeable with limited power consumption. In this paper, we propose a real-time and online tracking method to track multiple types of objects (e.g. pedestrian and car). Furthermore, to handle the ID switching problem, we provide a lightweight deep learning model which can recognize the similarity of objects. It can effectively solve the ID switching problem resulted from occlusion. Finally, we do some experiments to demonstrate that the proposed method achieves the state-of-the-art performance with less power consumption. The proposed method can solve the problem of high computation of tracking and keep the high accuracy of counting results with low ID switching rate. The experimental result shows that the average counting accuracy of the proposed method can reach more than 93% on pedestrian and vehicle counting applications. Also, it shows that the proposed method improves 68.2% on average of ID switching rate than previous works.