Component Detection And Its Application On Vehicle Detection

With the assistant of computer vision,the traffic information can be acquired directly by using of image capture system.As the key step of that system,the vehicle detection is a hot research issue in the area of Intelligent Transport System.However,in real-world problems,the detection of traffic object would pose a series of difficulties due to confounding variables such as illumination,partial occlusion and deformation.Thus,developing a robust object detection system is still a challenge.The component-based method has been recognized to be well suitable for object detection,and its idea is to represent object as an assembly of components and flexible spatial relations.Compared with global appearance-based method,component-based method could alleviate the difficulties of detection to some extent,especially for the issues about partial occlusion and deformation.Thus,the component-based method is a suitable solution for the detection for traffic object.The major purpose of this thesis is to explore some possible methods for promoting the performance of component-based method,thereby to facilitate the performance of final traffic information system.Functionally,the research of this thesis could be categorized as three layers.The first layer lies on the feature aspect,it aims to define a suitable feature that could integrate the information of components.The second layer,which is referred as the decision layer,mainly focuses on the information of components.The third layer mainly refer to the problem on the phase of detection,which aims to speed up the process of detection to satisfy the requirement of real-time.The mainly specific study can be list as following,1.To construct the component-fused features.Based on reviewing the wide references among this research area,a compact feature is found to have an ability to integrate the information of component,namely spatial histogram feature.The search for spatial templates in this feature could be considered as a process of seeking a series of suitable components for a specific object.Following this direction,two improved spatial histogram features are constructed.The first one base on the cues that obtained from the object itself.The basic idea is that to segment all positive instances in the training set,and then the final spatial templates are built referring to the centroids of clustering of the supperpixels.The constructing of second improved spatial histogram features is dependent on the accuracy of detection directly.Specifically,the process of searching for a suitable set of components is modeled as an optimization procedure.For solving this optimization problem,the evolutionary method is employed and the final accuracy of detection is used as the fitness function.2.The third research topic mainly focuses on algorithm.A new many-objective optimization method(crEA)is proposed,namely clustering-ranking evolutionary algorithm.The basic idea of this algorithm is to implement two operators(clustering and ranking)sequentially.Using a series of pre-defined reference lines as cluster centroid,the cluster operator associate each solution to a specific reference line via the projection distances.With reference to ranking operator,it rank the solutions that associated with each reference line according to the ability of convergence.Eventually,an environmental selection operation is performed on each cluster to promote both convergence and diversity for the final solutions.An extensive comparison with six state-of-the-art algorithms indicates that the proposed method is capable of finding a better approximated and distributed solutions set,especially in the many-objective optimization problems.3.A multi-objective clustering method based on crEA is proposed(MC-crEA).Two criteria of clustering,which reflect fundamentally different aspects of a good clustering solution,are optimized simultaneously.Considering the characteristic of vehicle detection,the locus-base adjacency scheme is employed as the representation for a solution.Besides,a sub-procedure for initialization based on two rounds of minimum spanning tree is proposed to alleviate the issue of over-convergence.In terms of identification of the best solution,a new model selection method based on final accuracy of system and a two-layers reference point system are proposed.Experimental results on synthetic and real datasets demonstrate the performance of the proposed algorithm.After that,the proposed method is employed for a vehicle detection task under the framework of bag-of-words.The results of experiment show that multi-objective clustering is able to promote the performance of codebook learning compared with single-objective clustering method.And the proposed MC-crEA is very competitive among the peer algorithms on the task of vehicle detection.4.An improved selective search method based on two new similarity measures.These two similarity measures are motivated from some basic image cues: saliency and normed gradients.In addition,an experimental investigation is expended for different objects under the traffic environment.It aims to provide a comprehensive solution based on different combination of similarity measures.