Research On Pedestrian Crossing Demand Detection Technology Based On YOLOv4

Pedestrian crossing is an important part of urban traffic optimization control.Accurate detection of pedestrian crossing demand is of great significance for ensuring pedestrian crossing safety at intersections and improving intersection traffic efficiency.Under the dual background of new epidemic prevention requirements and smart city construction,non-contact,safe,efficient and intelligent requirements are put forward for pedestrian crossing demand detection technology.Therefore,it is necessary to install non-contact video pedestrian crossing demand detection equipment in pedestrian waiting areas.The need is urgent.With the development of computer vision and deep learning,both deep learning algorithms and traditional machine learning algorithms for pedestrian detection have greatly improved the detection speed and accuracy.At present,video detection algorithms based on deep learning algorithms are in high computing power.The accuracy and real-time performance on the platform are already very high,but the realization of high-precision and fast detection on embedded chips with low computing power and small capacity remains to be studied.Based on YOLOv4,this paper conducts research on model cutting,network compression,algorithm optimization,etc,and finally realizes a high-precision and fast detection algorithm on an embedded chip with low computing power and small capacity,and develops an embedded pedestrian.Crossstreet demand detection equipment.The main work of the paper is as follows:First,pedestrian detection network acceleration research: In order to adapt to the low computing power and small capacity of the embedded platform,this paper makes lightweight improvements on the basis of the original YOLOv4.First,the network input image size is optimized according to the characteristics of the detected video to speed up the network operation.Then,according to the characteristics of the detection target,the large target detection network is trimmed to reduce the network size and adapt to the characteristics of embedded small capacity;finally,the calculation combination and quantification are used to improve the running speed of the network again.Finally,a relatively high-speed running speed was achieved on the Jetson Nano embedded module,and the frame rate was increased from the original 0.8FPS to 5.7FPS.Second,research on the accuracy optimization of pedestrian crossing algorithm:considering the influence of complex traffic scene environment,poor lighting conditions,outdoor rain and snow,etc.,first collect pedestrian crossing pictures in various scenes and different environments,and establish a special data set;secondly,for small pedestrian targets The feature of low resolution is enhanced by Mosaic data to increase the diversity of samples,enrich the background of the samples,and prevent the model from overfitting;finally,the training set is clustered by the K-means++algorithm to obtain the most suitable pedestrian features.The optimal aspect ratio and the number of anchor boxes reduce the convergence time of network training;Through the optimization of the algorithm in this paper,the accuracy of the final algorithm is improved by about 4.3% compared with the accuracy of the lightweight one.Third,the development of pedestrian crossing demand detection equipment.Based on the Jetson Nano embedded module,the overall framework of the industrial-grade pedestrian crossing demand detection equipment is designed,the circuit schematic diagram and printed circuit board PCB of the detection equipment are designed,and the debugging of the hardware prototype is completed;It runs on the hardware to verify the validity of the research results;the application software of the pedestrian detection equipment is compiled,and the prototype development of the video pedestrian crossing demand detection equipment based on low cost and high precision is finally completed.