An autonomous driving road surface target detection method based on adaptive fusion features

• The 3D sparse convolutional fusion network is designed to alleviate the loss of crucial information caused by variations in voxel feature dimensions, while simultaneously preserving a greater amount of spatial structural details within 3D scenes; • The regional suggestion network of deep and shallow feature fusion is studied to enhance feature expression by incorporating more detailed texture information into the deep feature map and capturing additional spatial geometry information in the shallow feature map; • The multi-task detection head network of attention mechanism is proposed to strengthen the weight value of important features and further improve detection accuracy The accurate detection of 3D targets on the road is a fundamental prerequisite for ensuring safe autonomous driving. This paper presents a novel single-stage approach for detecting 3D targets, which effectively tackles the challenges associated with extracting point cloud features from distant or occluded targets. Firstly, a 3D sparse convolutional fusion network is designed to effectively leverage multi-scale voxel features and enhance the expressive capacity of high-dimensional point cloud data in BEV view, thereby alleviating the loss of crucial information during voxel quantization. Secondly, the region proposal network incorporates deep and shallow feature fusion, where adaptive fusion is used to splice shallow texture features into the deep abstract feature map, thereby enhancing the expressive capability of target features. Lastly, the SE-attention mechanism is introduced in the multi-task detection stage as a means to enhance the network's accuracy in detecting, ultimately improving its overall performance. The experiments presented in this paper were conducted on the KITTI benchmark test, and the results demonstrate that our proposed method surpasses other 3D target detection techniques in detecting remote targets and handling occlusions. Specifically, we achieve a detection accuracy of 81.19% for cars and 65.32% for cyclists at moderate difficulty levels.