Research On Vehicle Detection And Tracking In Vehicle Active Safety

Advanced Driver Assistance Systems (ADAS) have been shown to effectively reducethe incidence of accidents and are a vital component of Intelligent Vehicle (IV). The accuracyand real-time performance of vehicle detection and tracking methods play a critical role in theoverall effectiveness of vision-based Forward Collision Avoidance Systems (FCAS) and BlindSpot Detection Systems (BSDS). Image-based vehicle detection methods using well-trainedclassifiers; such as those based on PCA+Neural Networks, Haar+Adaboost andHOG+Adaboost all require a scan of the whole image which involves time consumingcalculations making it impractical for real-time application. A process that has been widelyused in attempt to address this issue first involves the rapid identification of regions within animage that potentially contain a vehicle. This is done through the utilization of knowledge-,stereo-or motion-based vehicle detection methods and is typically referred to as HypothesisGeneration (HG). The next step in the process is referred to as Hypothesis Verification (HV).At this point, the candidate regions identified in the HG step undergo a process in which thepresence of a vehicle is either confirmed or denied. Confirmed vehicles are then tracked in thefinal step of the process; referred to as Vehicle Tracking (VT). VT is done by predicting thevehicle’s position utilizing information extracted from sequential images of the vehicle.Identifying the probable location of the vehicle circumvents the issue of “global scanning”and enhances the real-time capability of the technology.For single feature-based vehicle detection，We introduce five feature-based vehicledetection methods：shadow-based vehicle detection method, vehicle wave-based vehicledetection method，symmetry-based vehicle detection method, taillight-based vehicle detectionmethod and Active-Learning based HoG+AdaBoost vehicle classifiers. We first introduce themain steps of each feature-based method, and then these five methods are tested on the datasetof JVTL. Additionally, the main threshold setting process and performance of each methodare presented. Through the method comparison, the contribution of each method is clearlyshown to us.For multiple feature fusion vehicle detetion, after introducing the above singlefeature-based methods, we show how to use voting method, motion trajectory statistics,Choquet integral and D-S evidence theory to fuse these single feature-based vehicle detectionmethods. And then serveral videos which are captured from complicated environments areused to test these four methods. In order to verify the performance of these four methods, we use accuracy and algorithm processing time to measure these methods’ performance.The current vehicle tracking algorithms cannot meet the requirements of high robustnessin complicated engineering application scene. A co-training algorithm based on on-lineAdaBoost for vehicle tracking is proposed. First, a multi-vehicle tracking algorithm based onon-line AdaBoost and extended Kalman filter is used for vehicle tracking. Then, the trackingresult is verified by off-line classifiers which are learned from Haar feature and Adaboostalgorithm; this can handle the problems of vehicle exit and heavy occlusion between vehicles.Finally, the tracking window was reshaped according to the shadow of target vehicle.Experiments show that the proposed algorithm can handle the scale changing problem duiringthe tracking phase, has high robustness and flexibility with good application prospects.In order to improve the robustness of multi-vehicle tracking method, we present a novelmulti-vehicle tracking framework based on an improved particle filter. There are three majorcontributions in this method. First, a process dynamical distribution that has the capability toadapt to tracking multiple vehicles in complex environments is proposed in our particlefilter-based tracking framework. Second, we propose a target disappearance detection andhandling mechanism based on the normalised area of the minimum circumscribed rectangle ofparticle distributions (MCRP). This mechanism can be used for handling both thedisappearance of targets and determining whether a new target is a vehicle. Finally, wepropose an effective occlusion detection and handling mechanism which is based on theparticle interaction information. This was used to address the issue associated with occlusionduring the vehicle tracking process. Experimental results indicate that our method has theability to effectively and robustly track multiple vehicles in complicated road environments. Acomparison analysis is done by comparing our method with state-of-art methods and ourprevious work. The comparison analysis indicates that our algorithm is more reliable undervarious challenging conditions.