Research On On-site Verification Method Of Positioning Error Of Detection Systems For Vehicle Honking

The detection systems for vehicle honking combines the sound source localization system with the camera system to realize the visualization of the car whistle in the form of a sound phase cloud map,and it has been put into use in hundreds of cities across the country.There are only relatively preliminary reference opinions on the verification procedures,but there is no unified calibration or verification method to achieve quantitative detection of its positioning error.In this paper,according to the characteristics and principles of the acoustic cloud map,we combined the image detection and geometric derivation to calculate positioning error,conducted experimental research.And based on the image detection method,a positioning error detection algorithm based on deep learning was proposed.In addition,we used Python3.6 to configure the deep learning environment,then,the target detection algorithm based on YOLOX network and the semantic segmentation algorithm based on Deep Labv3+ network were trained respectively.The result is as follows:(1)The mean Average Precision of the YOLOX model reached 98.78%,the Mean Intersection over Union of the Deep Labv3+ model was 87.82%,and the Mean Pixel Accuracy was 93.85%,and achieved a good prediction effect.(2)The detection systems for vehicle honking were carried out on the outdoor road,and the positioning error was studied in depth according to different situations:static,dynamic,horizontal and vertical.The data of the dynamic and static experiments were input into the two network models,and it was found that the Deep Labv3+ network has a better detection effect.In the static experiment,the lateral positioning error of locating the sound source is less than 0.09 m and the longitudinal positioning error is less than 0.17 m.In the dynamic experiment,the lateral positioning error is less than 0.27 m and the longitudinal positioning error is less than 0.35 m.It can meet the requirements of on-site testing and has certain stability.(3)In the outdoor environment,through the Deep Labv3+ model,the static lateral positioning error of pseudo-color image of the detection systems for vehicle honking is less than 0.16 m,the static longitudinal positioning error is less than 0.39 m,the dynamic lateral positioning error is less than 0.35 m,and the dynamic vertical positioning error is less than 0.87 m.