Research On Vehicle And Pedestrian Detection Algorithm And Embedded System Application In Traffic Scene

With the advent of the 5G era,the development of artificial intelligence has entered a new stage,especially its application in the field of urban smart transportation construction,which has inspired scientists and created more room for improvement.Autonomous driving technology has always been the intelligent stage that human beings aspire to,but to achieve this great cause often requires the injection of infinite energy and wisdom.As an important research content in the field of autonomous driving,target detection uses various sensors to detect vehicles and pedestrians in complex traffic scenarios,providing more realistic and reliable road conditions and environment information for the improvement of autonomous driving technology.This thesis studies the detection of vehicles and pedestrians in actual traffic scenes.On the basis of ensuring the original detection accuracy,the detection speed is improved,and the real-time target detection in traffic scenes based on an embedded platform is realized.The main work contents are:(1)Yolov4 model lightweight processingThe Yolov4 model is selected as the basic framework for subsequent research,and the model is lightweight by replacing the CSPdarknet53 backbone network with the Mobile Net V2 network.Replacing ordinary 3×3 convolutions in PANet and Yolo head with depthwise separable convolutions reduces network parameters,reduces computation,and improves model detection speed.The Coord Attention mechanism is embedded in the Mobile Net V2 network to facilitate the model to more accurately locate and identify targets of interest.The volume of the Yolov4-mbv2-CA model after lightweight processing is 41.36 MB,and the detection accuracy reaches 83.78%,which is 5.86% lower than that of Yolov4.(2)Optimizing the model during trainingFirst,the dataset is expanded by mixing Mixup images and labels to balance positive and negative samples.Secondly,the Soft-DIOU-NMS post-processing method is used to replace the DIOU-NMS method that comes with Yolov4 to alleviate the problem of target occlusion in complex scenes.Finally,use the SGD optimizer to optimize the loss function to improve the detection effect in disguise.After the above optimization measures,the average detection accuracy of the final Yolov4-mbv2-CA model is 91.71%,which is 8.12% higher than the previous stage.(3)Establish STD data set to verify model effectThe real vehicle collects road traffic information in Beijing and creates an STD traffic data set.The experimental comparison shows that the Yolov4-mbv2-CA model has an average detection accuracy of 94.23% and a detection speed of 59 FPS on the STD data set.Compared with the Yolov4 model(95.32%),the detection accuracy is1.14% lower,which is almost negligible.The detection speed is 1.37 times that of Yolov4(43FPS).(4)Tensor RT model to accelerate processingAfter completing the construction of the embedded hardware platform Jetson Xavier NX and deploying the Tensor RT environment,the detection speed of the Yolov4-mbv2-CA model on the embedded platform Jetson Xavier NX is only 19 FPS,which cannot meet the real-time detection requirements.The Tensor RT optimized and accelerated Yolov4-mbv2-CA-trt(FP32+FP16)model increases the detection speed to75 FPS in the embedded system,which is about 3.95 times that of the Yolov4-mbv2-CA model before acceleration,which is enough to realize the real-time detection in actual scene.