Fast online multi-target multi-camera tracking for vehicles

Abstract

Multi-target multi-camera tracking (MTMCT) for vehicles, which aims to track multiple vehicles across multi-camera environments, is crucial in surveillance or intelligent transportation systems due to its broad applicability in real situations. However, the high inter-class similarity of vehicles and also their high intra-class variability due to the varying perspective, lighting, and video quality of each camera make it significantly challenging. Various offline approaches have been proposed and dominated the field with further advantages over the online strategy, but they are hardly adopted in real-world applications that usually require an online operation. In this paper, we propose a novel fast online MTMCT algorithm for vehicles considering better applicability in real applications. During the MTMCT, we actively reflect online MTSCT results, which is more reliable than clustering results in the multi-camera domain, on top of the object detection and feature extraction. To do so, we can effectively reduce the ID switches of the tracks and computational costs by decreasing the number of feature comparisons. As a result, we achieve 77.3 IDF1 on the S02 scenario of the CityFlow dataset with 0.012 seconds of tracking speed with four camera inputs. The source code is released at https://github.com/kamkyu94/Fast_Online_MTMCT.