Research On 3D Lidar Target Detection Algorithm For Autonomous Driving

With the rapid development of the communications industry and computer vision technology,autonomous driving has become a typical representative of the integration of science and technology and the real economy,and it plays a pivotal role in building the ecology of the automobile industry and promoting economic development.However,limited by the existing sensor hardware level and perception technology,the safety of the autonomous driving system needs to be improved.Therefore,the use of lidar point clouds for accurate three-dimensional perception has very important research value for the realization of high-level autonomous driving systems.Aiming at the shortcomings of existing point cloud target detection algorithms,this paper proposes a point cloud target detection algorithm that combines attention mechanism and adaptive adjustment to reduce information loss and overcome the characteristics of point cloud sparseness and uneven distribution.The innovative work of this article mainly includes the following two parts。1.Research on point cloud feature fusion algorithm based on attention mechanism.Aiming at the problem of information loss caused by the classic Point Pillars target detection algorithm during point-column segmentation,a more fine-grained method for generating multiple pseudo-images is proposed.Then,inspired by the basic idea of the attention mechanism,a feature fusion structure HC-Module with high attention and channel attention weighted in parallel was proposed to perform feature fusion on multiple pseudo images.The algorithm uses multiple weighting mechanisms to enhance the information expression capabilities of features.2.Research on point cloud target detection algorithm based on adaptive adjustment.Due to the working characteristics of lidar,the point cloud data collected by it has the characteristics of sparseness and uneven distribution,which increases the difficulty of the network to detect distant objects.To solve this problem,this paper proposes an adaptively adjusted target detection algorithm based on the above-mentioned feature fusion algorithm.This algorithm first enhances the feature extraction ability through the backbone network merged by the upper and lower levels,and fully excavates the hidden information contained in the feature map.Then,through the designed self-adjusting detection head,on the one hand,it overcomes the sparseness of the point cloud from the perspective of original information fusion,and on the other hand,from the perspective of adaptive adjustment,it reduces the interference caused by the uneven distribution of the point cloud to the detection.Finally,the modules are connected in series to solve the problem that the detection effect is limited by the point cloud characteristics.The experimental results show that under the same training data set,compared with the Point Pillars algorithm,the point cloud target detection algorithm proposed in this paper,which combines the feature fusion of the attention mechanism and adaptive adjustment,achieves higher detection accuracy.Especially for difficult samples,the algorithm in this paper has a better detection effect than many existing algorithms.