Research And Application Of Multi-sensor Calibration In Roadside Environment Perceptio

Roadside environment perception plays an important role in vehicle-infrastructure cooperation.Since the advantages of different sensors complement each other,the fusion of multi-modal data is more conducive to the comprehensive perception of road scenes.As the basis of various fusion algorithms,multi-sensor calibration results directly affect the performance of fusion.Unlike the vehicle side,traditional offline calibration methods in the roadside environment cannot correct the multi-sensor calibration deviation caused by specific factors in real-time.At the same time,due to the limited computing power of roadside edge computing units,calibration methods need to keep lightweight to reserve computing resources for other tasks.Because of the above characteristics,this paper combines target-based and targetless calibration methods to perform Li DAR-camera calibration in roadside environment perception.The main work and innovations are as follows:(1)A target-based(checkboard)Li DAR-camera calibration method is proposed.Due to the sparsity of point clouds,this method uses time-domain integration to extract features from multi-frame point clouds.According to the constraint of the reflection intensity distribution of point clouds on the checkerboard,the checkerboard measurement is used to estimate 3D corners.Combining 2D corners detected in the image,after corresponding corners in order,an external calibration matrix is generated by solving the nonlinear optimization problem.The experiments show that the method can perform accurate and stable calibration when the target is limited,and it can provide a good initial value for targetless online calibration methods.(2)A targetless Li DAR-camera calibration network(MCF-Calib Net)based on matching cost and calibration flow is proposed.The network is based on Calib Net,it introduces a cost volume in the feature matching layer to store the matching cost between RGB features and corresponding depth features.Calibration flow is introduced to loss function to predict the deviation between the initial projection position and ground truth to guide accurate correspondence of 2D-3D features.Finally,multi-range network iteration is used instead of single-range network iteration to further reduce calibration error.The experiments show that the network improves the accuracy of Li DAR-camera calibration,and it can correct random initial calibration deviations.(3)A roadside lightweight Li DAR-camera calibration network based on the depth map and NIN is proposed.The network is based on MCF-Calib Net for lightweight improvements.By performing effective stereo matching preprocessing,depth information is estimated from the image,and the multi-modal calibration process is reconstructed in the depth domain.In addition,multiple NIN structures are introduced to increase the depth of the network.The number of parameters is reduced while maintaining the same receptive field.The experiments show that the network improves the calibration efficiency under the premise of ensuring sufficient accuracy.