Research On Pedestrian Detection Method For Local Occlusion Scenarios

In recent years,object detection is a hotspot in the field of computer vision and artificial intelligence.As an important branch of object detection,pedestrian detection has been widely used in the fields of autonomous driving safety,intelligent security and intelligent robot.However,as a target of shape-shifting,the detection accuracy of pedestrian is not high due to various factors,such as pedestrians dress differently,occluded by pedestrians from each other or from other objects,diversification of pedestrian posture,diversity of scene background and different light intensity,which restricts the development of pedestrian detection technology.The main purpose of this paper is to conduct research on pedestrian detection methods,mainly studying the pedestrian detection methods under no occlusion and local occlusion conditions.After fully investigating the relevant contents of pedestrian detection,the main contents and contributions are as follows:Firstly,considering that the HOG feature does not have rotation and deformation and the internal texture information of pedestrians can't be well characterized,which leads to the unsatisfactory pedestrian detection effect.This paper proposes a pedestrian detection method based on the combination of HOG and block LBP feature(MULBP).The HOG feature describes the shape and contour information of the pedestrian,and the MULBP feature is used to describe the texture information of the pedestrian.The combination of the two can better represent the pedestrian information.However,the feature dimension after fusion is too high,and the generalization ability of detection is weakened.Therefore,PCA is introduced in this paper to reduce the dimension of the features after fusion,extract the essential expression information,remove the redundant information,and improve the detection effect.The comparative analysis experiment was conducted on the INRIA test set to verify the effectiveness of the feature fusion method based on PCA dimension reduction.Secondly,for the pedestrian images with partial occlusion,the original DPM detection method resulted in the deviation of the detection position or the target being missed due to the occlusion affects the score of the filter at the corresponding position In this paper,an improved DPM method based on probability density distribution was proposed.By constructing the occlusion model and the probability model,the proportion of the unoccluded filter is enlarged in the final objective function,and the proportion of the occluded filter is reduced,so as to adjust the score value of detection.A comparative experiment was conducted on ETH_IN_OCC of manually constructed occlusion data set to verify the validity of visible part model algorithm based on probability density function for local occlusion pedestrian image detection.Thirdly,in view of the low missed rate of small objects and occluded objects in SSD,this paper proposed an improved SSD pedestrian detection method based on context information combination,which introduces shallow features in SSD convolutional networks.The information is integrated with the deep semantic information and detected,and the pre-selection boxes with different aspect ratios are redesigned for pedestrian characteristics.Experiments in the occlusion data set show that the improved SSD model is better than the standard SSD model,and the improved SSD method can reduce the missed detection rate of local occlusion pedestrians and small target pedestrians.