Research On Intelligent Vehicle Positioning For Underground Parking Environment

As an important part of the advanced driving assistance system,the autonomous parking system can greatly reduce the parking accidents caused by the driver’s unskilled operation and other problems,and has a broad market prospect and research value.The positioning of the intelligent vehicle is an important prerequisite for the realization of the intelligent vehicle to complete the autonomous parking task,which is related to the realization of the subsequent autonomous decision-making,path planning and other important tasks of the intelligent vehicle.Underground parking lot as one of the main application scenarios of autonomous parking system,in this scenario,intelligent vehicle will face the problem of GPS failure,so visual inertial odometry becomes an effective choice for intelligent vehicle localization in this scenario,which can rely on only two low-cost sensors,camera and inertial measurement unit,to complete incremental localization of intelligent vehicle,and has been widely researched and developed.In this paper,we conduct an intensive study on intelligent vehicle localization in underground parking lot environment with respect to the characteristics of the scenario and the monocular visual inertial odometry method.Firstly,for the problems of low image brightness and uneven light distribution of images captured by cameras in underground parking lot scenes,this paper proposes a low-light image enhancement deep neural network based on Retinex theory and attention mechanism.The network mainly consists of three sub-networks-image decomposition network,illumination enhancement network and reflection denoising network.The image decomposition network is based on the Retinex theory and uses an encoding-decoding network to decompose the input image to obtain the illumination map and the reflectance map;the illumination enhancement network introduces dilated convolution instead of convolution-pooling to obtain a larger range of illumination map photometric information to enhance the illumination map;the reflectance denoising network introduces a CBAM attention mechanism module to obtain the attention map of illumination distribution in the illumination map to guide the reflectance map denoising.Finally,the enhanced illumination map and the denoised reflection map are used for reconstruction to obtain the final enhanced low-light image.Secondly,the underground parking lot has the characteristics of low texture and single texture,this paper proposes a visual inertial odometry method that incorporates monocular visual information and IMU information.The monocular visual inertial odometry method is mainly divided into two parts:front-end feature extraction and data association and backend nonlinear optimization.In the front-end,Harris corner detection is carried out for the image processed by the low-light enhancement algorithm,the pyramid layered LK optical flow algorithm is used for feature tracking,and the IMU data is pre-integrated;in the back-end,according to the front-end data association results,the back-end of the algorithm adopts the nonlinear optimization method of sliding window to optimize the visual reprojection error,IMU preintegration error and marginal information error in the sliding window,so as to estimate the pose information of intelligent vehicle.Then,to further improve the accuracy of intelligent vehicle pose estimation,a monocular visual inertial odometry cascade optimization module is proposed based on the pose of the keyframe marginalized by the sliding window and the map point coordinates associated with the keyframe.The module takes the keyframe pose and the coordinates of the map point associated with the keyframe as the initial values,and refines the keyframe pose by extracting and tracking the ORB feature points from consecutive keyframes using the motion-only bundle adjuestment to obtain a more accurate keyframe pose and generate 3D map points.Lastly,The proposed low-light image enhancement algorithm based on attention mechanism and Retinex theory and the cascade optimization visual inertial odometry method are combined to obtain the intelligent vehicle autonomous localization system in underground parking environment proposed in this paper,and the algorithm proposed in this paper is experimented and analyzed in detail by public data set and real vehicle offline experimental data.The experimental results show that the proposed low-light image enhancement algorithm can effectively improve the image quality;the proposed cascade optimization module can effectively improve the accuracy of monocular visual inertial odometry;the proposed intelligent vehicle autonomous localization algorithm in the underground parking environment outperforms the mainstream visual odometry and visual inertial odometry methods in terms of robustness and accuracy and achieves sufficient positioning accuracy,thus solving the positioning problem of intelligent vehicles in this scenario.