Research On Video-based Pedestrian Detection

Biological recognition technology has character of high reliability,good safety and strong robustness when it is used to identify human body features.Therefore,it has attracted more and more attention,research and application.Video-based biological recognition technology can automatically identify the identity of target personnel in video monitoring,and pedestrian detection is a hot research field in biological recognition.Pedestrian detection refers to the use of computer vision technology to determine whether there is pedestrian in the image or video sequence and give accurate positioning.It has been widely used in unmanned driving system,intelligent robot,intelligent video monitoring,human behavior analysis,and intelligent traffic and other fields.Compared with face detection,pedestrian detection is more difficult to detect pedestrian in various scenes due to the complex posture,larger deformation,more serious problems such as attachments and occlusion of human body.Aiming at pedestrian detection in video monitoring,this paper studies the traditional algorithm and the deep learning algorithm,respectively,and mainly improves the deep learning algorithm.The main work is as follows:(1)The traditional pedestrian detection algorithm is to design the human body's own appearance features manually,and then use the manual feature training classifier to identify the pedestrian and background.The first significant achievement is HOG+SVM detection algorithm,on which most subsequent traditional algorithms are improved.It is an important prerequisite for accurate pedestrian detection to design more discriminating and significant human features,also a classifier with good performance is an essential and important component.Therefore,this paper combines different feature extraction algorithms with different classifiers to study which combination method can achieve the best pedestrian detection effect.Experimental results show that the combination of aggregate channel feature(ACF)and AdaBoost classifier can achieve better performance in more complex outdoor scenes.(2)Many pedestrian detection algorithms based on deep learning have shown their effectiveness,but these algorithms are not accurate enough to locate the blocked pedestrian.Therefore,this paper proposes a convolutional neural network model(SCN)combining segmentation information and context information,which improves the accuracy of pedestrian rectangular box.The segmentation information sub-model of the SCN model separates the pedestrian from the background and then codes the score map through the LSTM network in order to reduce the impact of the occluded part on the location.Experimental results show that the SCN model is effective fordifferent degrees of pedestrian occlusion in different challenging data sets.(3)Due to the particularity of pedestrian detection application field,the requirement of detection speed is higher and higher.However,the pedestrian detection algorithm generally adopts a complex model with a large amount of computation,which is difficult to achieve real-time performance.Therefore,this paper proposes a new detection model to achieve a balance between accuracy and real-time performance.In order to adapt to pedestrians of different sizes,the pyramid strategy is adopted to extract pedestrian features,and the deconvolution layer is used to obtain higher context features.And the idea of deep supervision is introduced to guide the detection results to the ground truth.