Research On Perception System For Illumination Based Image And Spare Point Cloud In Autonomous Driving

Environment perception is one of the key technologies of autonomous driving system.The lack of visual features in low illumination scenes reduces the accuracy of image-based detection methods and makes it difficult to meet the perception requirements of autonomous driving.Therefore,aiming at the subject of low object detection accuracy in urban night scenes,this thesis proposes a night perception algorithm based on the fusion of image and sparse point cloud by introducing the prior knowledge of point cloud to improve network performance.The research mainly focuses on the three key technologies of "multi-modal cognition-attention mechanism-target-level fusion" of environmental perception system.Experiments show that the algorithm proposed in this thesis can effectively improve the object detection accuracy of autonomous driving in night scenes.The main contents are as follows:（1）Multi-modal feature fusion object detection method based on point cloud.Due to the loss of visual features,the accuracy of object detection is reduced.In this thesis,Proposed regions with high confidence are generated based on point cloud and given a higher weight in classification and regression tasks,so as to introduce an attention mechanism to make the network focus on the useful target features.In feature extraction,fusing multi-modal features of point cloud make up for the lack of visual features.Compared to the Faster RCNN detection network,mAP of this method improved by 2.3% under good illumination and 4.1% under low illumination,meeting the needs of object positioning of perception system,meeting the requirements of detection accuracy.（2）The method of object location based on the fusion of sparse point cloud and image.Considering that the sparse feature of point cloud is insufficient for object location,this thesis proposes a location method based on visual projection.The feature of image and lidar points are used for object detection.Instance segmentation and unsupervised clustering are used for point cloud detection.Object location is completed by integrating multi-source perception information based on visual projection.The detection accuracy of this method for objects of different classes is above 90% and the positioning error is less than 0.1m,meeting the requirements of target positioning of awareness system.（3）Establish of autonomous driving experiment platform and verification of algorithm.In the research,Hongqi EHS-3 is used as the basic platform to build an automatic driving system and a dataset covering multi-time campus scenes is constructed,which is applied to training and testing for network.In order to verify the effectiveness of the method,perception system proposed in the research is transplanted to the autonomous driving platform of Hongqi EHS-3 for road test based on real urban scenarios.The detection speed of the system increased from 7FPS to 13 FPS in the actual road test,meeting the real-time requirements of awareness system.This thesis introduces attention mechanism and multi-mode fusion technology to propose an object detection method based on point cloud,which effectively solves the problem of decreasing detection accuracy at night.An object location method based on the fusion of image and point cloud is proposed,which greatly improves the location accuracy.The method proposed can be transplanted to the autonomous driving platform to meet the requirements of multi-scene perception of autonomous driving.