Research On Unmanned Vehicle Environmentalawareness System On Unstructured Roads

With the development of science and technology,unmanned driving technology,due to its advantages such as intelligence and precision,has led to an increasing demand and research enthusiasm.Studying autonomous driving technology on unstructured roads can further enhance the ability of autonomous driving and enable its better application in military scenarios.The environmental awareness system,as the "eye" of unmanned vehicles to perceive the world,is a prerequisite and guarantee for smooth and safe driving of unmanned vehicles,and a key link in unmanned driving technology.Due to the advantages of Li DAR,such as being less affected by light,having a larger field of view angle,and precise 3D ranging,this article chooses Li DAR as the main sensor and focuses on studying the unmanned vehicle environment perception system based on Li DAR under unstructured roads.Firstly,this article simplifies the number of point clouds by statistical filtering and dividing regions of interest,and synchronizes the time of Li DAR and inertial navigation equipment.It compensates for point cloud distortion caused by unmanned vehicle bumps on non institutionalized roads using inertial navigation information to complete point cloud distortion removal.A new ground segmentation algorithm is proposed to address the insensitivity of existing ground segmentation algorithms to small obstacles in sparse point clouds.By calculating the overall point cloud concave envelope,the geometric boundary information of the 3D point cloud is fully utilized.The interior points corresponding to each concave triangle are proposed,and the point cloud category is determined based on the normal vector and position distance of the triangle.After experimental testing,the average accuracy of the algorithm reaches over 97%.Secondly,for the point cloud clustering task,this article uses depth maps for neighborhood point retrieval and analyzes the vector angle distribution characteristics with neighboring point clouds.Based on this,a judgment strategy is designed to not only classify adjacent objects,but also avoid clustering failures of tilted objects,making the clustering algorithm more robust and accurate.After experimental comparison,the algorithm in this article can cluster more targets,and the clustering within the cluster is more compact,and the clustering between clusters is more dispersed,The clustering effect is better.This article uses a multi target tracking module to associate cluster point clouds between different frames.Firstly,an interactive multi model filter is designed to predict the target position based on the different motion characteristics of the tracked target.Then,the target is filtered through the gated area and the target interconnection probability is calculated to complete the data association.Finally,in response to the problem of target recognition under unstructured roads,this article improves the Point Net algorithm based on convolutional neural networks.By associating each point between the two frame clustering point clouds associated with the multi target tracking module,the association distance is added to the point cloud as the original input feature,increasing the amount of information;Based on the depth map,the clustering point cloud is voxelated,and a 3D voxel convolution feature extraction module is added to enhance the network’s extraction of 3D shape features.In order to better verify the recognition performance of the network in this paper,experiments were conducted on various complex scenarios such as urban scenes,outdoor scene datasets,and self-test datasets.The recognition accuracy of the algorithm in this paper was higher than that of the comparison algorithm in most categories,and there was no significant fluctuation in the recognition accuracy for most categories.The average accuracy for each category of the network reached 73.87%.