Research On Pedestrian Detection Algorithm Based On Convolutional Neural Network

With the development of computer and various algorithms research,security monitoring and home care monitoring begin to apply computer vision technology more and more,and pedestrian detection becomes the first step of the function of gesture recognition and behavior analysis in monitoring applications.Currently,most popular pedestrian detection algorithms use convolutional neural networks and their derived neural networks to detect single-frame images,and perform poor performance in video sequence analysis applications.In this thesis,the security monitoring and home care monitoring needs oriented,for these common static background video sequence scenes,research and design pedestrian detection algorithm based on convolutional neural network to improve the performance of the algorithm.The main research contents are as follows:(1)Process INRIA data sets and expand it with related technologies that enhance data sets.Rich diversity of data sets to provide data support in order to improve the complexity of convolutional neural network and learn more network parameters.Common pedestrian detection datasets collect scattered pedestrians,background images or dynamic background video data shot by car-mounted cameras.In order to meet the requirement of static background video sequences,this thesis collects and marks some static background of pedestrian video sequences for pedestrian detection algorithm testing.(2)A pedestrian detection algorithm based on regional convolution neural network is constructed.By optimizing the network structure,applying regularization,DROPOUT and other techniques,the convolution neural network is trained repeatedly to obtain a better network convergence effect.Then,a large number of candidate regions extracted by the selective search algorithm are identified.The non-maximal value is used to remove overlapping and redundant candidate regions,and thepedestrian detection performance is better than the traditional HOG feature detection.(3)Aiming at the characteristics of the static background video sequence,the three-frame difference method is used to extract the candidate area,so that the dynamic target area in the video sequence can be efficiently extracted and the number of candidate areas can be reduced.The candidate regions are then supplemented using the Kalman filter tracking algorithm and the candidate regions are evaluated using the fine-tuned convolutional neural network.Eliminate accidental mismatch with SIFT-based inter-frame matching.Ultimately,the performance of the pedestrian detection algorithm for static background video sequences is improved.