Event-driven track management method for robust multi-vehicle tracking

Abstract

In this paper, we present an event-driven track management method to detect reliably and track robustly while minimizing missing and false detections. No state-of-the-art vehicle detection method can detect all the vehicles on the road without error. A multi-vehicle tracking method is essential to minimize the number of missing and false detections. In a multi-vehicle tracking method, there are three types of errors: false negative alarms, false positive alarms, and track identity switches. Our track management method can reduce the number of these errors remarkably while processing in real time for online application. Our track management method has four states: IDLE, PRE-TRACK, CUR-TRACK, and POST-TRACK. Most false positive alarms are removed in the PRE-TRACK state due to their sparseness. A track state transition to other states is determined by a track score. The track score is calculated by obstacle detection, vehicle recognition, detection-by-tracking, and data association. The proposed method is tested and verified with image sequences in real road environments. The experimental results demonstrate that the event-driven track management method minimizes the number of the false positive and false negative alarms remarkably compared with previous methods.978-1-4577-0891-6/11/$26.00 ©2011 IEEE

Looking at Vehicles on the Road: A Survey of Vision-Based Vehicle Detection, Tracking, and Behavior Analysis

“…In [127], extended Kalman filtering was used to estimate the ego motion, with independently moving objects' position and motion estimated using Kalman filtering. Vehicles were also tracked using extended Kalman filtering in [118].…”

“…In [118], vehicles were detected using an AdaBoost classifier on the monocular plane. The v-disparity was used to estimate the ground surface, and vehicles were tracked using extended Kalman filtering in the stereo-vision domain.…”

“…Integration of machine learning methodology could increase the robustness of existent stereovision approaches and has the potential to simplify the vehicle detection task. Research along these lines has been performed by using machine-learning-based detection on the monocular plane, integrating stereo vision for validation and tracking [5], [118], [126]. Future work could involve a more principled machine learning approach, learning on motion cues, image cues, and disparity or depth cues.…”

“…Estimation of the vehicle's yaw rate has emerged as a very important cue, for identifying turning behavior and for improved prediction of the vehicle's motion [82]. Extended Kalman filtering for vehicle tracking has increased in popularity to accommodate the nonlinear observation and motion models [82], [118]. Particle filtering has also increased in popularity for vehicle tracking while dispensing with some assumptions required for Kalman filtering [84].…”

“…Stereo-vision vehicle detection studies typically do not report true positive and false positive rates, although this has recently begun to change [118]. Instead, stereo-vision studies tend to focus on estimation accuracy for the motion parameters of a tracked vehicle, including position, velocity, and yaw rate [3], [84].…”