Pedestrian Detection Based On Deep Learning Algorithm Research

In the current booming computer vision discipline,pedestrian detection technology is a very important branch.In addition,pedestrians are the most frequently encountered targets in video surveillance and on-board camera shooting.The safety of pedestrians is the most concerned issue of social public safety.In order to avoid accidents and traffic accidents caused by crowded people,it is particularly important to detect pedestrians on the street in real time.For pedestrian detection in various practical scenarios,accuracy and real-time performance have always been key indicators.This article focuses on the difficulties of target detection in videos,and proposes a pedestrian detection algorithm that combines the ViBe algorithm and Faster R-CNN algorithm to solve the real-time problem of pedestrian detection in videos and improve the accuracy of detection.First,the foreground detection algorithm based on background update is studied,and the result of the first frame image detected by Faster R-CNN is used to initialize the ViBe algorithm model to avoid "ghosting".Then use Faster R-CNN's detection result of the previous frame to guide the update of the ViBe algorithm model in the current frame,and improve the robustness of the algorithm.Secondly,this paper uses the ViBe algorithm to extract the motion area to achieve the purpose of improving the calculation speed of Faster R-CNN and avoiding misjudgments.Then combined with the information in the time domain,the candidate regions extracted by the RPN in Faster R-CNN are screened according to the prior knowledge,and the detection time of the Faster R-CNN algorithm is further reduced.Re-study the convolutional neural network model.Aiming at the shortcomings of the traditional pooling algorithm of single feature extraction and insufficient flexibility,a cuckoo search pooling algorithm is proposed.The cuckoo search algorithm is introduced into the pooling algorithm,and the pooling parameters are continuously optimized until convergence.It has a certain degree of flexibility,and the improved algorithm can overcome the inherent shortcomings of the original pooling algorithm.Finally,a weighted time-domain softening non-maximum suppression algorithm is proposed.For pedestrian detection with overlapping parts in the image,the softened non-maximum suppression algorithm Soft-NMS is used to replace the traditional nonmaximum suppression algorithm NMS.Since the detection results of the same target in two consecutive frames may be completely different,the algorithm in this paper introduces time dimension information,and re-scores the classification box of the current frame by modeling the time domain results to ensure more accurate detection results.