Research And Application Of Pedestrian Detection Algorithm Based On Multi-feature Fusion

Pedestrian detection,as a key technology in the field of computer vision research,has a wide range of application values in the fields of intelligent monitoring,vehicle assisted driving,motion analysis and human-computer interaction.However,pedestrian detection is still a challenging problem due to the influence of complex scenes such as posture,dress,scale,light changes,and mutual occlusion.Fusion of multiple features can enhance the expression of human features and improve the performance of pedestrian detection algorithms.This dissertation focuses on issues such as multi-feature fusion,candidate region extraction,pedestrian scale diversity,and the application of pedestrian detection algorithms.The main research work is as follows:(1)Aiming at the pedestrian detection algorithms that combine histogram of oriented gradient(HOG)and Local Binary Pattern(LBP)features,the sliding window strategy is used to search when the scanning area is too large and the calculation is complex,resulting in slow detection speed,a pedestrian detection algorithm for locating target candidate regions is proposed.Firstly,a selective search algorithm is used to locate the target area,and the aspect ratio of the candidate area is limited to a certain range to filter out invalid windows.In view of the problem that the LBP operator only considers the difference symbol feature and oversimplifies the texture structure,the completed local binary pattern(CLBP)operator is introduced to enhance the expression ability of texture features,taking into account the influence of too high feature dimension of HOG and CLBP on the recognition ability of the classifier,the dimensionality reduction of the principal component analysis(PCA)is carried out and the series fusion is performed.Finally,the difficult sample mining process is introduced in the support vector machine(SVM)training to make the model training more fully and reduce the false detection rate.The simulation results on the INRIA data set show that the average miss rate of the proposed algorithm is 3% lower than that of the original algorithm,and the detection speed is increased by more than twice.(2)Aiming at the existing single shot multibox detector(SSD)network due to the large receptive field of deep features,the extracted target is relatively abstract and lacks detailed information,which leads to the problem of poor performance of small target pedestrian detection when applied to pedestrian detection,a multi-scale pedestrian detection framework that combines deconvolution to improve the SSD model is proposed.First,the deconvolution operation is used to upsample the deep features,introduce context information,and then merge with the shallow features for network prediction,a network layer containing Conv3＿3feature information is added during prediction to enhance the ability of small target pedestrian detection,network prediction layer increased from the sixth to the seventh.Aiming at the situation that part of the aspect ratio of the prior box generated by the multi-scale feature layer in the original SSD algorithm is not suitable for pedestrians,the aspect ratio of the prior box is redesigned,and the scale of this feature layer is set separately.The simulation experiment results show that the accuracy of the proposed algorithm on the VOC data set is improved by3.2% compared with the original algorithm,and the detection time of a single image is 0.028 s,the accuracy of the Caltech data set is improved by 2% compared with the original algorithm,the time is 0.032 s,which meets the real-time requirements.(3)In order to verify the applicability of the algorithm proposed in this paper in real scenarios,a pedestrian detection system is designed and implemented on the basis of the pedestrian detection algorithm proposed in this paper using the PyQt5 platform.The system consists of a login interface and a working interface,the working interface includes five modules: initialization,image acquisition,pedestrian detection,log recording,and visual interaction,the system uses the pedestrian detection algorithm proposed in Chapter 4 as the detection model to detect randomly collected pictures,offline video data and real-time video data.Experimental data in real scenes show that the system can effectively detect multiple pedestrian targets in pictures and videos,and meet the detection requirements in general scenarios.