Research On Unmanned Vehicle Target Detection And Recognition System Based On Sensor-fusion

In recent years,with the rapid development of automobile industry,driverless technology has been paid more and more attention by researchers.In the four modules of perception,decision-making,planning and control of unmanned driving,the perception part is the basis for the reliable operation and calculation of the subsequent modules.The sensing module obtains accurate environmental information through sensors distributed in various positions of the vehicle body.Because a single sensor has its own advantages and disadvantages,it is difficult to obtain comprehensive environmental information using a single sensor.The resulting target detection algorithm based on sensor fusion can well combine the advantages of each sensor,so as to achieve the purpose of obtaining accurate and comprehensive environmental information.It is favored in the future because it can provide accurate information for driverless vehicles;Camera is widely used because of its low price and conducive to target recognition.By fusing the sensing results of lidar and camera,the advantages and disadvantages of the two sensors can be complementary,so as to obtain comprehensive environmental information.Therefore,this paper proposes a target detection method based on the fusion of lidar and camera to realize the detection and recognition of road targets.The specific research contents are as follows:(1)Research on target detection based on image recognition.In this paper,YOLOv4 neural network model is used as the model of image recognition;The model training adopts the mixed method of Pascal VOC data set and self built data set,which mainly improves the proportion of each label category and realizes the accurate recognition of specific road targets.In addition,a monocular ranging model based on camera is established to extract the distance of obstacles in front.The experimental results show that YOLOv4 model can meet the requirements of intelligent driving in terms of detection accuracy and speed.(2)Research on target detection based on lidar.Firstly,the original point cloud data is filtered and downsampled,and then the ground point cloud is segmented by using the improved RANSAC algorithm,which can not only segment the point cloud on flat ground,but also identify the ramp.For the point cloud after ground segmentation,the improved DBSCAN(density based spatial clustering of applications with noise)clustering algorithm is used.The clustering threshold is changed from a fixed threshold to a variable distance threshold that changes with the distance of obstacles.The experimental results show that the clustering algorithm has significantly improved the segmentation accuracy and over detection rate.(3)Joint calibration of lidar and camera.Firstly,this paper uses the joint calibration toolkit based on the framework of autoware to calibrate the sensor.Because the corresponding feature point pairs need to be selected manually in the calibration process,artificial point selection is an important reason for the calibration error.This paper uses the improved joint calibration algorithm to change the extraction of feature point pairs from artificial designation to automatic selection,which reduces the calibration error and improves the calibration speed.(4)Research on target detection and recognition based on data fusion.The obstacle detection results(clustering results)of lidar are projected onto the image to form their respective detection areas with the recognition results of the camera.Based on the idea of multi hypothesis,the correlation probability is constructed to correlate the detection results of the two sensors.The fusion range model is established for the correlated results to extract the accurate spatial information of the target.