Empirical Analysis of Vehicle Tracking Algorithms for Extracting Integral Trajectories from Consecutive Videos

Abstract

This study introduces a novel methodological frame-work for extracting integral vehicle trajectories from several consecutive pictures automatically. The frame-work contains camera observation, eliminating image distortions, video stabilising, stitching images, identify-ing vehicles and tracking vehicles. Observation videos of four sections in South Fengtai Road, Nanjing, Jiangsu Province, China are taken as a case study to validate the framework. As key points, six typical tracking algorithms, including boosting, CSRT, KCF, median flow, MIL and MOSSE, are compared in terms of tracking reliability, operational time, random access memory (RAM) usage and data accuracy. Main impact factors taken into con-sideration involve vehicle colours, zebra lines, lane lines, lamps, guide boards and image stitching seams. Based on empirical analysis, it is found that MOSSE requires the least operational time and RAM usage, whereas CSRT presents the best tracking reliability. In addition, all tracking algorithms produce reliable vehicle trajecto-ry and speed data if vehicles are tracked steadily.