Detection Algorithm Of Road Falling Objects Based On Fusion Of Visual Image And LiDAR Point Cloud

Roads are an important economic lifeline in our country.With the increasing vehicle load on the road in recent years,the amount of road falling objects is also rising.Road falling objects are a huge hidden danger of causing traffic congestion and traffic accidents.Therefore,it is very important to detect road falling objects accurately and quickly.At present,there are few research reports on road falling objects detection systems and algorithms,and the existing methods cannot fulfill the demand of detection accuracy.Aiming at the road falling objects detection,this paper analyzes the characteristics of road falling objects,conducts in-depth research on the hardware layout and detection classification algorithm of the road falling objects detection system,and designs a falling objects detection scheme based on visual image and LiDAR point cloud fusion and a falling objects detection algorithm FE-SSD.Specific work includes the following:(1)The overall scheme of the road falling objects detection system is proposed,and the sensor selection and system layout design of the falling objects detection are completed,including four modules.The data acquisition module uses environmental perception sensors such as LiDAR,CMOS industrial camera,GPS positioning system,and inertial measurement unit(IMU)to collect road information and vehicle information;the data processing module uses a vehicle-mounted industrial computer to realize data processing and storage functions;The data display module selects the onboard display to output the processing results;the power supply module provides electricity for the entire system.(2)A new algorithm for detecting falling objects using LiDAR and cameras is proposed.The advantages of high positioning accuracy of point cloud processing and high classification accuracy of image processing are respectively used to improve the accuracy of falling objects detection and optimize the detection rate and false detection rate of falling objects detection.The algorithm includes two steps of pavement target extraction and falling objects classification.First,the road edge is detected through the point cloud image,and then the ground point cloud filtering and point cloud clustering are performed to extract the target point cloud cluster within the road surface range;then the target point cloud cluster is projected into the visual image aligned in time and space to obtain the corresponding image region of interest;then use the improved Res Net-50 network for object classification.Finally,the prediction information of road scattered objects is obtained.(3)An algorithm FE-SSD for small object detection is proposed.First,Darknet-53,which has good accuracy and efficiency,is used as the backbone network,and the group residual boosting algorithm is used to utilize the ability of channel features;then the receptive field is expanded by paralleling multiple convolutions of different scales to improve the sensitivity of multi-scale features.learning ability;in addition,use ECAM to strengthen the important channel information of the target;finally,a feature fusion algorithm based on ECAM is proposed.It mixes the semantic information in the deep features into the shallow layer,and achieves a deeper fusion.(4)An experimental system for road falling objects detection was built with an electric scooter as a carrier,data was collected in the campus environment,and experiments were carried out on the image and point cloud fusion algorithm and the FE-SSD algorithm.The average detection precision of the image and point cloud fusion algorithm is 94.84%,the recall rate is 91.92%,and the detection effect is good.However,the detection ability of small-sized falling objects is poor.The average precision of the FE-SSD algorithm in the VOC small target dataset is 67.5%;for falling objects objects with a size less than 10 cm,the detection precision of FE-SSD is 73.33% and the recall rate is 64.71%,which makes up for the deficiency of the image and point cloud fusion algorithm proposed in this paper in the detection of small-size road falling objects.