Vehicle Trajectory Extraction And Behavior Analysis In Complicated Traffic Video Scence

Vehicle trajectory extraction and behavior analysis based on video is an interdisciplinary research field,covering the technologies of digital image processing?artificial intelligence and pattern recognition.However,due to the complexity of study object in this field,there are many difficult problems to be solved.Moving vehicle detection,tracking and behavior analysis based on video in complicated traffic environment have been the research difficulties and hot spots in this field,many methods and technologies are not perfect enough.In this dissertation,some key technologies such as vehicle detection,moving vehicle tracking,similarity measurement of vehicle trajectory and trajectory clustering are studied deeply,and the main contribution are as follows:1)For vehicle detection in complicated traffic environment,a vehicle detection method based on vehicle symmetrical and shadow features is proposed.Based on SURF algorithm,a novel symmetrical transformation is presented to translate a nonsymmetrical SURF descriptor into a symmetrical one.Because of the larger matching errors with symmetrical points distributed at different scales,the S-SURF algorithm is improved and optimized by reducing the accumulation times of Haar features and minimizing the impact of scale on feature points.Then the optimized S-SURF algorithm are used to extract vehicle symmetrical features,which will be further used to locate the central position of vehicle.Finally the target vehicle will be recognized with the shadow features under it.In this method,the symmetrical SURF and shadow features of vehicle are used for vehicle detection and recognition.The main advantage of the proposed method is the avoidance of vehicle segmentation from complex background.2)Stable and reliable vehicle tracking is a key step of vehicle trajectory extraction.A novel vehicle tracking method is proposed by combining feature matching and optical flow.Based on the bidirectional reversibility constraint KLT algorithm,a new offset estimation method is constructed,which will delete feature points with poor stability,consequently the reliability and stability of feature tracking is significantly improved.Meanwhile,SURF algorithm is used as a compensation mechanism for updating and adjusting the feature point sets.Then,according to the relative location and relative angle of feature points in the first frame,the scale and rotation variation of objects in the current frame is determined.Finally,hierarchical clustering is used to delete abnormal feature points,and the object region is determined in the current frame.With the fusion of two matching strategies,this algorithm can not only improve the stability of tracking algorithm,but also solve the problems of object deformation and partial occlusion.In addition,it is robust and effective to scale and rotation variation3)Trajectory similarity measurement is an essential question in the process of trajectory clustering.Because of the complexity and diversity of vehicle trajectory,the existing measurement method has its limitations.A trajectory similarity measurement method is proposed by incorporating multi-features to edit distance.Based on EDR,the speed and direction features of the trajectory are used to segment the trajectories,and different edit operation values are assigned to segments with different feature meanings.In the end,the IEDR algorithm based on segmentation is further defined and analyzed.The proposed similarity measurement not only has the benefits of EDR,such as allowing timewarp and noise immunity,but also further improves the accuracy and robustness of trajectory similarity measurement.4)With the purpose of extracting the normal motion mode in traffic scene,vehicle behavior mode learning is the precondition of abnormal vehicle behavior recognition.A trajectory cluster model using Bayesian maximum a posterior estimation based on the incremental DPMM is proposed.In this dissertation,each vehicle trajectory is represented by a DFT coefficient vector,and a DPMM-based method is presented for trajectory clustering.Then,an incremental version of DPMM-based trajectory clustering algorithm by improving Gibbs sampling process is further developed.The conditional distribution used for Gibbs sampling is designed by fixing the states of the previous trajectories and updating the states of new trajectories.Meanwhile,the normal moving mode of trajectory is learned to identify abnormal behavior of vehicles.While the DPMM is learned,an appropriate number of trajectory clusters is automatically determined without training data,and the DPMM model can adaptively change with model parameters and classification numbers,which can solve the training difficulty under the data sparsity due to unpredictability and uncommon of abnormal behaviors.In addition,all the new trajectories will be assigned to the existing trajectory clusters or to new trajectory clusters with the help of existing clusters,and the cluster efficiency is greatly improved as there is no need to cluster all the trajectories each time.