Research On The Fault-tolerance Perception Methods Of Intelligent Vehicle Based On Multi-information Fusion

It is essential for Intelligent Vehicles(?)to perceive the environment accurately.The existing schemes realize accurate perception based on the multi-sensors,which ensure the hardware redundancy of the system.However,these plans increase the complexity of calibration because of too many sensors.At the same time,a large number of data will put forward the high requirements of the real-time of the vehicle processor.In addition,the high cost of hardware is not helpful for the mass production and application of the sensor schemes.Therefore,the study of multi-sensors perception algorithms and models are of great significance to enhance the ability of intelligent vehicles fault tolerance.This paper focuses on the tasks of fault-tolerance perception,and studies the virtual vision generation,depth estimation,depth completion and 3D object detection.From the perspective of practical application,the feasibility of perception technology in intelligent vehicles based on multi-information fusion is discussed.The specific works are carried out as follow:1.A virtual vision generation algorithm based on generative adversarial networks is proposed to solve the problem of binocular vision system which cannot work stable when the lack of vision signal happens.The reason of vision signal missing is occlusion,pollution and other extreme conditions.This algorithm introduces 3D convolution and attention mechanism to extract the time-series feature from image data,depth data and optical flow data,which improve the robust of sparse features.Meanwhile the discrimination network is introduced to realize the confrontation training between models,which further reduces the reconstruction error of virtual vision.Finally,the experiments show that the proposed algorithm can improve the accuracy and the inferring speed of virtual vision significantly under extreme conditions,and further verify the practicability in the trajectory tracking tasks based on the SLAM system2.A monocular depth estimation algorithm based on graph convolution neural network is proposed to solve the problem that the distance measurement is unfeasible when the vision signal is missing.Different from the traditional algorithm which needs to match the same pixel position to obtain the disparity information,the convolutional neural network is employed to get the depth information of the monocular image,and get rid of the calibration distortion,signal missing,signal pollution and other problems caused by extreme conditions.This algorithm introduces the pixel-level depth topological graph and graph neural network to obtain the depth correlation feature and to avoid the gradient vanish problem in deep networks.In addition,the loss function of depth topology structure and multi-scale structure are employed in the model.The former makes up the sparsity of depth topological features,and they both further improve the accuracy of depth estimation.Compared with the traditional methods,this algorithm has a significant advantage in estimating the depth of small volume objects.3.In order to solve the low accuracy of perception according to the sparse data from low beam LiDAR,a depth completion algorithm based on sparse point cloud is proposed.The algorithm is main to generate the dense depth information by the information fusion of sparse point cloud and image.Firstly,asymmetric convolution is used to extract image feature to get rid of the feature imbalance.Secondly,the attention mechanism is used to decouple the image features in frequency domain to ensure the independence between different frequency features.Finally,the depth completion results are obtained by feature fusion.The experiment results show that the proposed algorithm can reduce the complexity of the model,improve the density of the sparse point cloud as well as the perception ability.4.In view of the problem that the intelligent driving scheme lacks redundant substitution in the perception when the signal of high-beams LiDAR is missing,an 3D object detection framework based on pseudo LiDAR is proposed.In this algorithm,the sparse point clouds collected by low-beams LiDAR and the images collected by camera are used to generate the virtual point clouds.And the depth correction module is applied in the network to improve the accuracy of virtual point clouds.According to the improved RPN network,the imbalance feature problem is solved and the accuracy of proposal bounding boxes are improved.The experiment results show that the performance of the proposed 3D object detection algorithm based on pseudo LiDAR is similar to the high-beam LiDAR.