Research On Pedestrian Detection In Real Scene Based On Convolutional Neural Network

Pedestrian detection is to determine whether the input image or video contains pedestrians and accurately find out the specific location of the pedestrian.As a sub direction of object detection,pedestrian detection is has a wide range of applications in video surveillance,pedestrian recognition,image retrieval,and advanced driver assistance systems.Because pedestrians have non-rigid attributes,it determines that pedestrian detection is different from ordinary object detection.In addition,there are many factors that restrict pedestrian detection,such as the complex diversity of backgrounds in real-life scenes,changes in lighting,pedestrian occlusion,changes in posture,and diverse shooting angles,real-time requirements,small target pedestrians,etc.These factors have brought great challenges to pedestrian detection.Therefore,pedestrian detection has always been a research hotspot and difficulty in the field of computer vision.In this paper,in order to improve the detection performance of pedestrians and small-scale pedestrians in real-world complex scenarios,the methods based on the traditional manual features and the methods based on deep convolutional neural networks are studied respectively.The main researches are as follows:(1)Aiming at the problems of pedestrian detection methods in real scenes such as missed detection,high false detection,and low detection accuracy of small-size targets,a pedestrian detection(PDIS)algorithm based on an improved SSD deep network model is proposed.The algorithm improves the original SSD network model by eliciting lower-level output feature maps,and uses the abstract features of different layers of convolutional neural network to detect pedestrian targets separately,fuses multi-layer detection results,and improves detection performance of small-target pedestrians..The PDIS algorithm trained on the augmented pedestrian data set improves the accuracy of pedestrian detection in real scenes.Experiments show that the PDIS algorithm achieves an accuracy rate of 93.8%on the INRIA test set,and the missed detection rate is as low as 7.4%.(2)The diversity of dataset samples can effectively improve the generalization ability of detection algorithms.This paper has collected pedestrian images under different lighting,pose,occlusion and other complex scenes,and has expanded the more complex INRIA pedestrian datasets with more backgrounds.The set has been posted on github(https://github.com/csust7zhangjm/CSUSTPD).CSUSTPD contains samples from different scenes such as schools,streets,and stations,which are combined into a real-life pedestrian data set in a complex background.The gestures,occlusion,and small target pedestrians in the training sample are all marked,such as:riding Bicycles,umbrellas,crowded pedestrians,etc.CSUSTPD complicates and diversifies the sample background of pedestrian data sets,and greatly increases the number of small target pedestrian annotations.The PDIS algorithm trained on the augmented pedestrian data sets improves the accuracy of pedestrian detection in real scenes.(3)For the problem of high misdetection rate of ACF pedestrian detector and poor performance evaluation of DeepLabV2 algorithm,this paper combines ACF algorithm and DeepLabV2 algorithm to improve the performance of pedestrian detection.Firstly,the pedestrian is detected by the ACF detector and all detected bounding boxes are saved.Then the coordinate information of all bounding boxes is mapped to the semantically-segmented output picture,and semantically segmented pedestrian pixels are searched in each frame mapping region.The proportion of occupied pixels is determined to be a pedestrian when the ratio of the percentage of pixels occupied by the pedestrian pixels in each bounding box is greater than a certain threshold.Otherwise,the pedestrian window is detected as a false check and deleted.Experimental results show that compared with the original ACF detection results,the fusion algorithm of ACF and DeepLabV2 reduces the false detection rate of pedestrians detected by the ACF algorithm and increases the robustness of pedestrian detection.