Research On Moving Target Recognition Based On Lidar Point Cloud Depth Completion

With the advent of the fourth industrial revolution,computer and artificial intelligence technologies have been rapidly developed,based on that as an emerging industry autonomous driving has also received a lot of attention from academia and industry.Compared with traditional driving modes,a high-performance driving decision-making system can avoid various traffic accidents caused by drivers,free human hands and relieve traffic pressure at the same time.Accurate and sensitive perception algorithms are the prerequisite for autonomous driving of unmanned vehicles.However,the present hardware redundancy solutions require high hardware costs.As the consequence,the increase of sensors will also bring great challenges to data synchronization and algorithm real-time performance.Therefore,it is of great significance to research on low-cost and strong anti-interference perception algorithm for the mass production of unmanned vehicles.This paper takes the precise target perception of unmanned driving as the research target and conducts in-depth research on the perception problems to be solved urgently such as motion blur,depth completion,and 3D target detection,afterwards the information fusion methods of different sensors are discussed.The main work and results of this paper are as follows:1)Aiming at the problem of motion blur in the image collected by the camera due to vehicle vibration and relative motion in the driverless scene,this paper proposes an image deblurring algorithm MobileNet-GAN.Since the depth separable convolution used in MobileNet can save massive computing resources compared with traditional convolution,we used it as the backbone structure.At the same time,in order to reduce the number of parameters of the entire model,we share the weights of nonlinear layers in different scales,which greatly reduces the inference time.The comparative experiments results on multiple datasets show that the inference speed of our model surpasses other comparison algorithms,which can basically meet the preliminary real-time requirements in the driverless scene.Finally,we designed a Mask-RCNN target detection network based on ResNet,which proved the practical value of the MobileNet-GAN algorithm.2)In order to solve the problem of sparse point cloud information caused by low beam LiDAR used by unmanned vehicles,a multi-modal feature fusion depth completion algorithm based on point cloud and image is proposed.In view of the problem that the information from remote positions needs to be transmitted through a multi-layer network when the partial convolution is used to process the sparse point cloud,our algorithm introduces the gated convolution in the global channel to obtain the feature dependence in a larger area.To reduce the inference time of point cloud depth completion network,this paper uses ERFNet and Stacked Hourglass module to predict the surface normal vector and the depth of local channel,which greatly improves the efficiency of our algorithm.Through comparative experiments,the algorithm proposed in this paper is verified on the standard depth completion dataset KITTI and achieves higher inference speed and accuracy3)Aiming at the shortcomings of existing target detection network in 3D point cloud,a moving target detection algorithm is proposed based on the voxelization operation of point cloud.In this paper,spatial attention mechanism and channel attention mechanism are introduced to update the feature information of key points to enhance the dependence of peripheral information.In order to solve the problem that the special structure in the local area will be ignored in the process of key point selection,which leads to the decline of the detection effect of small targets and long-distance targets,this paper introduces a deformable convolution module after 3D sparse convolution to extract more distinguishing feature information.The proposed detection algorithm is verified on KITTI target detection dataset.The experimental results show that our algorithm achieves the optimal detection effect in pedestrian and bicycle categories,and maintains a high level in the vehicle detection.