Reserarch And Implementation Of Pedestrian Target Detection Based On Video Stream

Pedestrian target detection is one of the important research branches of computer vision,which has very important application significance for pedestrian detection.Such as in the fields of unmanned driving systems,human-computer interaction systems and intelligent security systems.With the continuous advancement of technical means,the feature extraction has gradually developed from the initial manual construction of features to an end-to-end network in which features are extracted and classified by a deep neural network.As far as the current method is concerned,good results have been achieved for pedestrian target detection in simple scenarios,but the ideal results have not been achieved in the actual application environment.There are two reasons:First,due to the influence of natural factors such as light and angle.Second,because the human body is a non-rigid structure,it is prone to occlusion.These two situations will lead to the lack of pedestrian feature information and make the classifier misjudge,making the implementation effect poor.Therefore,in response to the above two problems,this thesis proposes to use video stream as input,combined with the advantages of timing information,and proposes a method for constructing inter-frame timing information features.The model of this method aggregates the convolutional features of the detection frames to avoid Natural factors and occlusion make the pedestrian feature information missing,and miss and misdetect the situation,thereby improving the accuracy rate and other evaluation indicators.And based on this pedestrian target detection algorithm,a pedestrian detection system is designed and implemented.1)This thesis is a pedestrian detection application based on video streams,using Resnet^[28]for feature extraction,and studying a method of pedestrian target detection based on deep learning.The main frame uses Faster R-CNN^[21]network structure for detection.It is conducive to the process of feature aggregation and processing,and realizes an end-to-end network structure,which simplifies training and deployment.2)The input of pedestrian target detection in practical applications is mostly video stream input.Video stream input has more information than static image input.In order to better utilize this advantage,a feature aggregation method based on time-series detection frames is proposed to make pedestrian feature information more abundant and complete,and avoid blurring and blurring of a certain frame of static image.Information is missing,pedestrian targets are too small,and pedestrians are blocked to affect the detection effect,which are the unique advantages of video popular people detection.The detection result shows that the method in this thesis also has a higher anti-interference ability in actual scene applications,and the accuracy of detection results.3)Use public data sets to verify the algorithm,and on this basis,use pytorch,My SQL and Py Qt to design and implement an intelligent monitoring platform to realize multi-screen function,and at the same time bear the video stream image input of the camera to send to the server for detection And the function of receipt result solves the online and offline dual pedestrian target detection function.