A Study On Pedestrian Behaviors Detection Based On Deep Learning

With the development and progress of people s living standards,the number of vehicles in China has shown a rapid growth trend.Vehicles bring great convenience to people s life and work.However,the traffic accidents caused by vehicles have also increased,resulting in a large number of people and property loss.According to authoritative statistics,three quarters of dead people caused by traffic accidents in China are those have weak traffic right,and the lack of protection for them in tensifies this situation.Therefore,for the protection of those people,avoiding the occurrence of traffic accidents has always been the focus of researchers.Observing and judging the actions of pedestrians and cyclists on the road,providing a basis for subsequent decision-making planning and other steps,real-time warning to reduce the probability of traffic accidents is a direction of current research.Based on the video extraction image,this paper studies the main problems of current vision-based pedestrian behavior.The main research work has three parts: sample collection cleaning and behavior labeling,image preprocessing and pedestrian behavior detection and recognition.Firstly,the pedestrian behavior data under different environmental conditions,such as strong and weak illumination,backlight,rain fog and occlusion,are collected.Instead of using a single photo,the image is extracted from the video frame as a training sample to achieve a single frame for real-time imaging of the camera.The dynamic blur effect of the image is as close as possible to the pedestrian movement in the real scene.For some relatively rare actions,data enhancement is used to increase the sample size to prevent under-fitting.Then,for the problem of illumination and image blur,mainly through the filtering method,firstly use the limit contrast adaptive histogram equalization method to alleviate the situation that the target is not obvious under the light illumination dark or rain fog environment,and enhance the ability to resist the interference of illumination changes;The median filtering method in smoothing filter weakens the image noise under low light conditions and improves the image quality.Finally,the Laplacian filtering method in gradient filtering is used to sharpen the image and enhance the edge information of the object.Images that are not high image blur have some improvement.Mark all behavioral samples as seven categories after completing the above steps.Finally,based on the idea of YOLOv3 deep learning detection algorithm,darknet is selected as the basic neural network structure,different hyperparameters are selected to train the samples,and the migration learning method is used to compare the effect of the trained model,and the optimal training model is selected for real-time detection.