Research On Unmanned Object Detection Method Based On LiDAR Point Cloud And Image Fusion

With the rapid development of the economy,the ownership of private vehicles is booming increased.The increasing number of ownership makes traffic congestion and accident more frequent,resulting in huge loss of lives and economics to individuals and the country.The emergence of unmanned driving technology provides a new direction to mitigate this situation of traffic congestion and frequent traffic accidents.Object detection is one of the main tasks of unmanned driving,existing researches focus on object detection with single sensor,which could not adapt to complex and variable road scenes.Therefore,to make full use of information collected by sensors to ensure the object detection model be more adaptive.Employing multi sensors and data fusion method to detect object in unmanned driving is becoming a hotspot.In addition,it can also be extended to the corresponding target detection problem in the electrical field to achieve practical application in the electrical field.This paper focuses on object detection in unmanned driving scenarios,and conducts research on key technologies such as image-based,point cloud-based,and image-point cloud fusion-based object detection.The main work is as follows:(1)A object detection model based on multi-scale feature fusion(GA-YOLO)is proposed.Images possess rich semantic information and have unique advantages for object detection,after an extensive review of existing image-based detection methods.Based on YOLOv5 s,this paper firstly introduces dilation convolution and designs a spatial pyramid module(SPP-Atrous)based on dilation convolution for replacing the multi-level pooling operation in the original spatial pyramid pooling to avoid the occurrence of gradient vanishing.Secondly,to simplify the structure of the model,a lightweight multi-scale feature fusion backbone network(GS-CSPNet)based on Ghost is designed in this paper to enhance the multi-scale feature extraction capability and detection efficiency of the model.(2)A detection algorithm based on hybrid attention difference point cloud(HA-RCNN)is designed.The point cloud data has depth information that is not available in images,and the LIDAR is more resistant to interference and has better detection capability at long distances.Therefore,this paper designs the HA-RCNN object detection algorithm based on point cloud data.To address the problem of large loss of foreground points in the sampling process of current object detection algorithms,a hybrid sampling module(HS)is designed in this paper to make the distribution of sampling points more balanced and avoid large loss of foreground points.In order to fully extract point-wise features,this paper introduces a self-attentive mechanism and designs a hybrid attention module(HA).After experimental verification,the algorithm is able to accurately differentiate the foreground points from the background points in the point cloud,achieving higher road object detection capability.(3)A detection algorithm based on image and point cloud fusion(LRNet)is constructed.The image data has semantic information and the point cloud data has depth information,and the two types of data can be fused to achieve complementary information.Therefore,in order to improve the reliability of object detection,the LRNet detection algorithm is proposed in this paper based on a decision-level fusion scheme.The algorithm is based on the image as well as point cloud based object detection algorithm proposed in this paper,encodes the detection results of both detection methods,fuses the detection results in the LR-Fusion module and finally filters the generated prediction frames by Soft-NMS.The fused detection network enables parallel detection method,which greatly improves the reliability and safety of unmanned environment perception.At the same time,the fused detection method also demonstrates superior detection performance for small targets such as pedestrians.