Road Multi-Objective Real-Time Detection Method Based On Improved YOLOv3 Algorithm

In a complex and varied road traffic environment,the driver's negligence and lack of concentration can cause irreversible damage.During the driving process of the vehicle,realtime classification and positioning of surrounding objects is a key part to ensure road traffic safety.Aiming at the problem that the existing road target detection technology is difficult to accurately classify and locate multi-target objects in real-time in video images,this paper introduces the real-time classification and localization of multi-object objects on the road by target detection.The work includes the following aspects:(1)The development process of fully connected network,convolutional network and deep residual neural network is introduced.The basic components of convolutional neural network are described in detail(including low-dimensional convolution,multi-dimensional and multinuclear convolution and the working mechanism of 1×1 convolution kernel,etc.)and the calculation process of deep neural network feature forward propagation and error back propagation.(2)The design principle of the feature extraction network composition of the YOLOv3 algorithm and the target detection feature map output network is analyzed.The calculation process of the target object bounding box prediction regression and the design mechanism of the loss function are elaborated.(3)Aiming at the characteristics of the boundary frame of the road target object,based on the Yolov3 algorithm,an improved YOLOv3 deep residual convolutional neural network architecture model with five feature detection maps and 155 layers is designed.The 7×7 and 104×104 output detection maps are used to detect larger and smaller target objects in the road traffic field of view,respectively.The pictures of the network training comes from the BDD100 K dataset.The experimental results show that compared with the YOLOv3 architecture,the improved multi-detection maps YOLOv3 network has an average accuracy of 54.48% on the verification data set,which is an increase of 5.11 percentage points.(4)For the distance measurement of objects,two methods based on least squares curve fitting and longitudinal distance measurement based on camera focal length are tried.The experimental results show that the latter's target longitudinal distance measurement error is smaller(The average error is 4.05%.)At the end of this paper,the target object detection of the improved YOLOv3 algorithm and distance measurement are combined,and the function of outputting the detected object information in the program is modified,the program outputs 6 categories of target object classification,position information and the longitudinal distance between the 4 categories of target objects and the camera,and the video detection speed reaches 29.8 frames per second,which meets the requirements of real-time performance.This method can provide reference for car-assisted driving in natural road traffic scenarios.