Research On 3D Object Detection Algorithm Based On LiDAR Point Cloud

Three-dimensional(3D)object detection is a long-standing research task in the field of autonomous driving.In autonomous driving scenarios,the LIDAR sensors can produce point cloud data accurately,so it has become one of the most frequently used point cloud collection devices.Because of the sparsity and density inhomogeneity of point cloud data,the LIDAR-based 3D point cloud object detection is still full of challenges.Therefore,towards autonomous driving scenarios,this paper proposes a series of 3D object detection algorithms based on LIDAR point cloud.The main work and innovation of this paper are as follows:(1)Due to the sparsity and density inhomogeneity of point clouds,the current methods are still struggling to provide rich information for the objects with different scales,leading to lower performance.To solve the problem,this paper takes the PointRCNN two-phase network as the baseline,and further presents a 3D point cloud objects detection algorithm based on cross self-attention mechanism,named CSA-RCNN.Firstly,a cross self-attention based point cloud feature extraction network is proposed,which constructs a multi-scale space by KNNs with different radius.Meanwhile,this network uses a cross-attention to capture the contextual information from each scale space.Secondly,a multi-scale fusion strategy is designed to aggregate the contextual information from different scale spaces,thus improving the information extraction ability for the objects with different scales.Finally,a bin-based regression loss function is introduced,it generates a higher accuracy initial3 D wraparound box.The experimental results show that the CSA-RCNN significantly surpasses the baseline network in terms of detection accuracy,especially for the objects with small-size.(2)Existing methods usually extract local features within the RoI via the ensemble abstraction,ignoring the learning of the correlation between the sampled key points.Therefore,these methods fail to effectively recognize the position and orientation of the objects.To address this problem,with the presented CSA-RCNN algorithm,this paper designs a CSA-SE detection algorithm based on RoI enhancement.First,an overlapping sampling strategy is proposed to sample the extended RoI for obtaining the key points with the multi-density distribution.This strategy can enhance the saliency of local features within the RoI and alleviate the noise caused by CSA-RCNN to expand the prediction envelope box.In addition,a squeeze-stimulated attention mechanism is constructed,to model the correlation between the sampled key point.It can improve the ability to aggregate the local features within the RoI,such that the location and orientation of objects can be recognized effectively.Extensive experiments demonstrate that the proposed algorithm predicts the position and orientation of objects more accurately,compared with existing mainstream methods.(3)This paper explores the deployment problem of the proposed algorithm on intelligent edge computing platforms,which further promotes the application value of 3D target detection in autonomous driving scenarios.Specifically,by TensorRT acceleration optimization and network model transformation,this paper implements the deployment of the CSA-SE algorithm on the intelligent edge computing platform NVIDIA Jetson TX2.Experimental results illustrate that the proposed can migrate to smart edge computing platforms as well as maintain an excellent detection performance.