Research On Algorithm And Application Of Pedestrian Detection Based On YOLOv3

The intersections of urban roads are highly prone to traffic accidents,accounting for about 30% of the total.The accidents at night intersections account for a small percentage of the total,but the injuries are more serious.Affected by the light at night,car drivers may not be able to observe pedestrians at the intersection ahead in time.They can inform automobile drivers of the pedestrian information through the camera at the intersection.The car drivers can take timely measures to avoid pedestrians,which can guarantee the safety of pedestrians.This thises mainly studies the application of the YOLOv3 algorithm to pedestrian detection at intersections.According to the characteristics of the detection object,YOLOv3 is improved to improve the detection accuracy of the model.At the same time,YOLOv3 is ported to TX2 which an embedded platform.Aiming at the problems of TX2 computing power and limited memory,the model is pruned,compressed and accelerated to improve the detection speed of the model.Finally,the communication between TX2 and V2 X equipment is established.When TX2 detects pedestrians at the intersection,the pedestrian information and location are sent to the vehicle V2 X equipment to achieve pedestrian warning effect.The main research work and results are as follows:(1)Taking pedestrians at night intersections as the research object,integrating the performance of various algorithms,then YOLOv3 is selected as the basic model of pedestrian detection;owing to the preset anchor size in the original YOLOv3 cannot meet the task requirements of pedestrian detection,k-means++ algorithm is used to re-cluster anchor size;improved bounding box regression function to solve the problem of inaccurate positioning;contrary to the characteristics of the pedestrian detection data,structure and parameters of the network is adjusted.Experiments result shows that the accuracy of the improved pedestrian detection model is 4.5% higher than that of the original YOLOv3 model,and the detection speed also meets the needs of real-time pedestrian detection tasks;(2)In order to port the model to the embedded platform TX2,the pedestrian detection model was pruned and fine-tuned.The experiment proves that the improvement the volume of the model after puring and fine-tuning is reduced to 1/8 of the original,and the reasoning speed is increased by 3 times.Use TX2 acceleration library Tensor RT to accelerate the pedestrian detection model,build an acceleration framework,and the pedestrian detection model was integrated with the framework.The experiment results shows that the detection speed of the model is improved by 18 FPS when the detection accuracy lost by 1%;(3)Apedestrian warning system is constructed at intersections based on TX2: TX2 is responsible for transmitting the detected pedestrian information to JS-road embedded roadside devices through serial port communication;JS-Road uses the DSRC communication protocol to package the pedestrian information and the GPS information in the module.Vehicle-mounted embedded equipment JS-car;through the data processing module of JS-car,calculates the distance between JS-car and JS-road,and then displays the pedestrian and distance information through a human-computer interaction interface to achieve the function of pedestrian warning.Finally,the experiment proves that the system designed in this thises can realize the pedestrian warning function.