Research On 3D Object Detection Algorithm Based On Gaussian Kernel Function Heat Map And Fusion Information Distance

The rise of autonomous driving has garnered significant attention from the academic and industrial communities towards three-dimensional object detection and autonomous driving technology.The current mainstream approach for threedimensional object detection is based on point cloud data.However,due to the sparsity,limitations,and complexity of point cloud data,despite the existence of many excellent algorithms,there are still a series of issues such as low robustness and poor generalization.In light of the research conducted in this paper,the focus is on the theoretical algorithm research of three-dimensional object detection based on point cloud data and related detection models.There are still pressing problems in current methods that need to be addressed,including:(1)Most existing three-dimensional object detection algorithms based on point cloud data predict the object’s center point and perform box regression through neural network operations to obtain the predicted bounding box for the object to be detected.However,since all the point cloud data is mainly located on the surface of objects and the labels for the centers are very small and similar,training the object’s center points from point cloud data is theoretically unreasonable.This approach performs poorly in prediction and distinguishing from other points,often resulting in false detections.(2)Most existing three-dimensional object detection algorithms based on point cloud data use the farthest point sampling algorithm based on spatial information distance during point sampling.This leads to the omission of many important foreground points before the final bounding box regression,thus affecting detection performance.In particular,the detection performance of small objects such as pedestrians is fundamentally limited,severely impacting the model’s accuracy.To address problem(1),where existing methods overlook the peculiarity of point cloud data,this paper proposes an algorithm based on Gaussian kernel heatmap to predict the corner points of object bounding boxes.Firstly,drawing upon the algorithmic ideas from two papers in the field of 2D object detection,namely Corner Net and Extreme Net,the proposed approach utilizes Gaussian kernel heatmap to obtain the corner points of the predicted object bounding boxes and their corresponding information,thereby establishing initial predictions.This algorithmic approach better aligns with the unique characteristics of point cloud data,leading to improved robustness and accuracy of the algorithm model.To address problem(2),where existing methods discard a large number of foreground points,leading to poor detection performance for small objects like pedestrians,this paper proposes a new point sampling criterion that incorporates fusion distance,which combines spatial information distance with feature information distance.By employing the fusion distance-based farthest point sampling algorithm,most foreground points can be retained,compensating for the deficiencies of solely relying on spatial information distance in the previous farthest point sampling algorithm.Furthermore,while preserving an adequate number of foreground points to enhance regression accuracy,a sufficient number of background points are also retained to improve classification accuracy.To validate the effectiveness of the proposed algorithm,extensive experiments were conducted.The experimental results on the KITTI dataset demonstrate that the proposed improved algorithm in this paper achieves higher classification accuracy for cars,pedestrians,and cyclists compared to existing methods.Additionally,the visualization results indicate that the proposed algorithm can more accurately predict object bounding box information.