Research On Multiple Object Detection Model From The Roadside Perspective Based On Deep Learning

With the development of wireless communication technology,establishing the Intelligent Vehicle Infrastructure Cooperative Systems（IVICS）has become an inevitable development trend to empower autonomous driving in the fields of over-thehorizon perception and efficient interaction and collaboration of traffic participants.As an important part of the Internet of Vehicles,the Roadside Sensing System relies on the edge computing platform with low computing power to form structured data,realize accurate and real-time sharing of over-the-horizon perception information,and solve problems such as the limitation of single-vehicle line-of-sight.It has become an urgent need for research.This paper combines the research on the visual information-based object detection technology of the roadside perception system in the key research and development plan project of Hebei Province.At present,in terms of object detection algorithms applied in the field of transportation,traditional methods have gradually declined due to their complex implementation methods and inefficient calculation processes.Object detection algorithms based on deep learning have replaced the above methods and have achieved certain research results.However,the current research status usually focuses on improving the detection accuracy and efficiency from the single-vehicle perspective,ignoring the synergy between the roadside system and the vehicle.There is still a lack of research on public datasets of traffic participant detection algorithms from roadside perspectives and multiple object detection models that match the characteristics of roadside perspective image information.Due to the above research background and current situation,this paper builds a multiple object detection model for traffic scene information from a roadside perspective based on the deep learning method.The main contents are as follows:1.A vehicle object detection dataset from a roadside perspective is constructed.In order to obtain training and testing datasets suitable for the application scenarios and requirements of this paper,this paper analyzes the current mainstream public datasets in the field of autonomous driving,summarizes the production specifications and labeling formats of such data sets,and builds an autonomous test data collection platform based on the theoretical guidance of public data sets.Collect multiple object data information from the roadside perspective on the test road section,perform image preprocessing to standardize image information and correct distortion,and use labeling tools to manually label multiple types of vehicle targets.According to the practical application requirements of this paper,the RS-UA dataset with roadside perspective characteristics is obtained by fusing the public dataset and the self-collected image data.2.A lightweight multiple object basic detection model based on deep learning is constructed.This paper analyzes the deep learning object detection network model of the current mainstream one-stage detection model YOLO and the lightweight model Mobile Net,establishes Mobile Net V3 instead of CSPDarknet53 as the backbone network of the detection model,and reduces the params of feature extraction network through depth separable convolution;SPP processes the output feature map of the deep network,selects the output of different depth feature maps of the lightweight backbone network,and fuses the deep semantic information and shallow apparent information through the PANet to form the neck of the detection model;in the detection model head in this paper,three network outputs with different feature map sizes are set up,so that the object information of different sizes of the same image can be obtained to the object information at the appropriate network depth,and the construction of the lightweight detection model M3-YOLOv4 is completed.Through model training and testing,the m AP of M3-YOLOv4 on the dataset RS-UA is 0.906,which is 1.1% lower than the value of the one-stage detection model YOLOv4,and the params of M3-YOLOv4 model is reduced to 10% of YOLOv4.The model’s forward inference time under the same platform also has obvious advantages.3.A one-stage lightweight multiple object detection model is constructed from a roadside perspective.This paper analyzes the characteristics of spatial invariance of image information from the roadside perspective,proposes an optimization method based on the deep learning object detection model,and improves the attention mechanism used by the Inverted Residuals in the original feature extraction network.The optimization method of coordinate attention mechanism enhances the long-range dependence of neural network on spatial information.The traditional convolution operation in the attention mechanism is optimized by the Depthwise Overparameterized Convolution method,which improves the network’s ability to extract and express vehicle features in the high-dimensional feature space.The models trained and analyzed in this paper are compared and tested.The optimized one-stage lightweight multiple object detection model DCM3-YOLOv4 in the RS-UA dataset has a m AP value of 0.930 and a network model params of 31.12 M.The inference time is 96.13 ms under the CPU,which is superior to the basic model YOLOv4 on the same platform.DCM3-YOLOv4 is used to conduct field tests,and online detection of vehicle objects in traffic scenes from the roadside perspective is carried out.After the test,this model can realize real-time detection under CPU.In summary,the multiple object detection model based on deep learning proposed in this paper has strong applicability and advantages in roadside perspective application scenarios.