Research On Preceding Vehicle Tracking Algorithms For Driving Assistance

The Advanced Driver Assistant System(ADAS)refers to the vehicle equipment that uses advanced technologies such as sensor,computer vision and communication technology to reduce the incidence of active accidents.In ADAS,the most important procedure is the perception of the driving environment,where the vehicles are the main obstacles when driving.The real-time perception of the vehicle information can effectively maintain the safety distance between the vehicles and prevent from car collision.With the rapid development of computer vision technology,the institutions all around the world are contributing their major force for the vision based vehicle tracking technique.Vehicle tracking plays the key role of locating the target vehicle.Different from vehicle detection,the vehicle tracking can utilize the associated information of the neighbor frames in the video sequence to predict the possible position of the target at the next time,thus avoiding the global search of the video sequence to ensure the real time performance of the algorithm.However,there are still many challenges in the vehicle tracking task because of the complexity of the background information,the uncertainty of the target vehicle movement and the low calculation performance in the actual driving environment.At present,there are mainly two kinds of tracking algorithms,which is generative tracking and discriminative tracking algorithm respectively.In this paper,we first investigate the application of these two algorithms in vehicle tracking,and then introduce a real-time and robust tracking algorithm for vehicle tracking.First,because the tracking target is pre-defined,we can analyze the intrinsic features of the vehicles for tracking.Through the comparison of the color,texture and edge features of different expressions about the vehicle,this paper proposes a new SULBP feature for tracking,which combines Sobel operator and LBP.This new feature can not only preserve the texture information of the object's internal structure,but also the Sobel description of the high-frequency information of the edge of the object.It also has the insensitivity to illumination.Indicated by the experiment results,this new SULBP feature outperforms other features for vehicle tracking.Second,in order to improve the time performance of the tracking algorithm,the dimension reduction strategy is required for features.In this paper,following the philosophy of the Uniform Local Binary Pattern(ULBP),we proposed an adaptive dimension reduction method,which calculates the statistical binary-feature transition.Then we applied it to the SULBP feature.Indicated by the experiment results,this strategy can reduce the running time of the algorithm to a great extent.Third,considering the single feature is not enough for the expression of the target,we proposed a multi-feature fusion particle filter algorithm.Through the analysis of the distribution of the particle set between the color feature and the SULBP feature under different interference,we introduced a fusion strategy Based on area of the minimum circumscribed rectangle of particle set.Indicated by the experiments,this multi-feature fusion particle filter algorithm based on area of the minimum circumscribed rectangle of particle set(MCRP_PF)outperforms the single feature particle filter and the fusion coefficients fixed particle filter algorithm,which is more robust and has better time performance.Finally,based on the Kernelized Correlation Filters(KCF)algorithm,we introduced a particle filter based scale evaluator to solve the drawback of KCF for adaptive scale evaluation.The new scale adaptive KCF algorithm(AS_KCF)outputs the location and scale of the target through traditional KCF and particle filter scale evaluator respectively.Indicated by the experiments results,this AS_KCF algorithm outperforms some of the current excellent discriminative tracking algorithm,with mean success rate increased by 12.3% compared with traditional KCF.Meanwhile,this algorithm is demonstrated can be fully applied for real-time application on PC.