Orientation-based Feature Aggregation for Multi-Target Multi-Camera Tracking

Abstract

Multi-target multi-camera tracking (MTMCT) is a crucial component in traffic flow analysis for smart transportation systems. MTMCT generates vehicle trajectories from the surveillance videos across the cameras at different intersections. Variable vehicle orientations during the driving process degrade the MTMCT precision. Besides, time-varying traffic flow between cameras could lead to trajectory mismatching under constant travel time constraint algorithms. In this paper, an orientation-based MTMCT method considering time-varying traffic flow is proposed. First, stacked re-identification backbones (SRB) are employed to merge the features extracted by different backbones for generating a discriminative appearance feature. Second, to re-duce the information loss, orientation-based feature aggregation (OFA) is proposed to represent a trajectory consisting of vehicles with different orientations comprehensively. Third, considering the time-varying traffic flow, dynamic spatio-temporal strategy (DSS) is designed to segment the whole videos into time windows and construct fine-grained travel time probability functions to narrow the gap between the model and realistic situation. Experiments validate the effectiveness of the proposed MTMCT method. Our method outperforms other works on the IDP metric.