A Real-time Distributed Multi-camera Multi-object Tracking System

Abstract

Multi-camera Multi-object Tracking (MCMOT) is a challenging problem. Most of the existing methods use a centralized architecture to achieve high tracking accuracy. However, as the number of devices increases, the communication and computing needs of a centralized architecture will grow superlinearly, making it difficult to use in city-scale scenarios. In this paper, we propose a real-time distributed MCMOT system, Uni-ID, which ensures that the communication and computing costs of each device remain almost constant as the number of devices increases. Uni-ID uses the existing single camera detection and tracking architecture and adds a feature extraction network between the detection and tracking phases. By broadcasting the extracted features, the device can reach a consensus on the object so that an object has the same ID in different devices. The system has been tested with real-device recorded data and achieved state-of-the-art performance