Research On Roadside Lidar Data Sensing Method Based On Multi-algorithm Fusion

With the sustainable development of intelligent transportation and technology of connected and automated vehicles,advanced sensing sensors as an important tool in the evolution of intelligence and information technology have been widely deployed in all corners of urban transportation.Lidar,as a sensor that can accurately detect the location of objects,has strong environmental adaptability,strong anti-optical interference ability,and high spatial resolution.It has gradually become an important way of intelligent monitoring.However,at present,lidar is mainly installed at the vehicle end,and limited detection is realized with the help of high-performance on-board computers.For the deployment of lidar at the road end,and how to reduce the demand for computing resources,the research of using traditional algorithms to achieve effective detection is not enough.Based on the above information,this paper uses roadside lidar to carry out research on target perception.The main contributions are described as follows:(1)According to lidar characteristics and equipment requirements,the roadside data perception platform was designed,and the data acquisition scene was investigated.The erection height and inclination Angle of lidar were analyzed and verified from theoretical calculation,numerical simulation and scene experiment respectively,and the optimal erection scheme of roadside lidar for complex intersection scene was finally determined.(2)According to the characteristics of the collected point cloud data,the data set establishment scheme is developed.Firstly,the unqualified point clouds are deleted through artificial processing,and the qualified point clouds are classified,and the special point clouds are saved in advance.Secondly,the key frame is extracted from the data set by the interframe difference method to obtain the data set that needs annotation processing.Finally,the category of point cloud annotation is set,and the PCAT after secondary development is used to annotate it,and the roadside lidar point cloud data set used in this paper is obtained,and compared with the existing roadside data set.(3)Based on the distinguishing features of the background point cloud and the existing point cloud background filtering algorithm,a background filtering algorithm based on history frames is proposed.In this method,the history frames are processed and merged.Then the density filtering is performed using the inconsistency of the density of the background point cloud and the foreground point cloud.Finally,K-D tree is used for background differential filtering.Experimental results show that the proposed algorithm can achieve 98% extraction rate of background point cloud and 96% extraction rate of foreground point cloud.(4)For the issue of target detection in the context of sparse point clouds,this paper proposes the distance-based DBSCAN clustering method to perform cluster analysis on point cloud clusters.In addition to ensuring the applicability of the point cloud classification method and the reasonableness of the parameter settings,this paper improved the AOA optimization algorithm,and used the improved ITAOA algorithm to optimize the SVM algorithm.Finally,the accuracy of target detection is compared between the traditional algorithm and the three deep learning algorithms used in this paper.The experimental results demonstrate that the traditional algorithm can achieve the same accuracy as the deep learning algorithm which requires high computational resources in the range of 0-60 m,under the condition that the traditional algorithm does not need the support of higher computing resources and maintains real-time performance.In summary,this paper has completed the design and construction of the roadside data perception platform,designed and analyzed the lidar deployment scheme,and proposed a new point cloud background filtering algorithm,which realizes the research on the target perception method of point cloud data based on traditional algorithms independent of high-performance computers,and has practical significance for the further deployment and landing of the roadside lidar.