Research On Key Technologies Of Pedestrian Detection In Images

Driven by applications of driverless car,visual surveillance,human-robot interaction,et al.,detection,recognition,and tracking specific objects in the environment have become a hot research area in computer vision.In the environment,one of the most concerned objects is pedestrian.In recent years,the research on pedestrian detection is developing rapidly and has preliminarily been applied in industry.The core of pedestrian detection is a binary classification problem,in which the difficulty lies in the high variability of background and pedestrians' appearance caused by different scales,occlusion,and so on.To tackle these variances,this thesis studies feature extraction and classification algorithm.The contributions of the thesis are summarized as follows.Firstly,this thesis proposes feature selected self-similarity(FSSS)features.The proposed method is based on self-similarity(SS)features.To solve the high dimensionality problem of SS,Linear Discriminant Analysis(LDA)is used to select features with good property of inter-class discrimination and intra-class invariance from all the SS features,which reduces the dimensionality of SS features.The method is with high efficiency,easy to extend to other detection tasks,and the features selected are intuitive and reasonable.Experiments show that the detector trained with FSSS features achieves 13.96% Log Average Miss Rate(AMR)on Caltech pedestrian dataset,which is better than other non-deep learning methods.Secondly,the thesis proposes Scale Aware Pooling(SAP)and Soft Decision Tree(SDT)to detect multiscale pedestrians.The proposed methods are based on Multiresolution Filtered Channels(MRFC).To solve the problem that features' receptive fields are not correspondent for pedestrians of different scales,SAP is proposed to make the receptive field of a feature change with the pedestrian scales.To relieve the problem that features are not scale invariant,SDT is proposed to tackle pedestrians in different scales in a divide and conquer way.Experiments show that by combining these two methods and two efficient sliding windows classification strategies,the trained detector achieves 13.84% AMR on Caltech pedestrian dataset,which is better than other non-deep learning methods and with a detection speed of 20.15 FPS.Thirdly,this thesis proposes Hierarchical Multi-Pose Learning(HMPL)to tackle the occlusion problem.The method is based on Multi-Pose Learning(MPL).To solve the problem of tackling too many occlusion patterns separately,a hierarchical structure composed of cells and parts is proposed.This structure takes advantage of the characteristic that occlusion patterns overlap each other and enable part detectors to share decision trees.This algorithm automatically trains part detectors for different occlusion patterns when the occlusion patterns of training samples are unknown and lower the weights of the occluded regions,which lessens the interference to the classifier caused by these regions.Experiments show that the detector trained by HMPL achieves 62.87% AMR on Caltech pedestrian dataset when the pedestrians are heavily occluded,which is better than detector trained with ordinary Boosted Decision Trees(BDT).In summary,the thesis researches key technologies of pedestrian detection with the perspectives of feature extraction and classification algorithm.Through optimization of SS features,MRFC and MPL,methods to tackle the variance of background and pedestrians are enriched in this thesis.The research result supplies references for the improvement of accuracy and speed in pedestrian detection systems.