Objective Laser range-gated 3D imaging is a new type of 3D imaging technology for long-distance detection. Fog, rain, snow and other severe weather conditions have been regarded as one of the technical challenges that hinder the landing of autonomous driving in recent years. This technology has the characteristics of suppressing backscattering and increasing the effective distance. At the same time, it can achieve 3D imaging of the target with millions of pixels, showing great potential for long-distance detection in severe weather such as fog, rain, and snow. Traditional gated 3D imaging methods have problems such as high system complexity, dependence on hardware characteristics, poor system flexibility, and difficulty in balancing accuracy and real-time performance. The existing visual guidance method does not consider the visual characteristics of the gated slice image, resulting in limited accuracy. Affected by the rear radiation, the traditional RGB camera effectively detects very low in the dense fog and strong light environment. Although scanning laser radar can obtain accurate distance information, it is limited by mechanical scanning angle, resulting in low space resolution of long -distance detection timing; They are difficult to meet the long -distance detection and perception needs of autonomous driving under bad weather conditions.
Methods We proposed a vision-guided range-gated 3D imaging method that integrates an attention mechanism. Starting from the visual level, this method focuses on calculating regional weights for object contours, areas with weak textures, and other areas to improve regional prediction accuracy. A lidar depth completion algorithm is combined with the true value used for model supervision to obtain a dense depth truth image, thereby further improving the model's depth estimation accuracy.
Results and Discussions The results are shown in Fig.6. In the figure, it can be seen from the rectangular box area that the proposed method has made a clear qualitative comparison with other methods. The first line is an enlarged display of the rectangular box content. The proposed method has higher detection accuracy and precision for objects at a longer distance, such as cars, traffic signs, and trees at night, while keeping the edge details of the target clearer. From the figure, it can be seen that the imaging effect of the SGM model is poor, there are regional imaging errors, low imaging accuracy, and many image noise points. The results of the methods compared with the AdaBins model, Monodepth model, and PackNet model have been improved. The imaging effect is better at a distance, but the details around the object are omitted. Overall, the imaging effect of the target is still poor. In contrast, the Gated2depth model provides a clearer imaging effect, better imaging accuracy for the target, but there are erroneous imaging of the target edge, and low imaging accuracy for the distant area. Compared with other methods, the proposed method has higher imaging accuracy in the target contour, and good imaging effect in the open area and weak texture area in the distant area. It can be seen from the comparison around the vehicle in the enlarged area that the proposed method effectively reduces the interference of noise, making the overall imaging accuracy higher. The results of the comparative test evaluation indicators are shown in Tab.1-2 below, where G2d represents the abbreviation of Gated2depth.
Conclusions The result has higher depth prediction accuracy than existing methods. The experimental results show that it has higher accuracy in object contours. Through comparative experiments in bad weather (rain and snow), it is more resistant to interference than other methods, and has better range-gated imaging effects. It has the ability to suppress noise in the image and has strong processing capabilities for weak textures, obtaining clearer target edges and retaining more details. The MAE in night data increased by 6.3%, and the effectiveness of the proposed method was verified by experiments. In the future, it is necessary to further improve the network based on the optical characteristics of the gated image to improve the accuracy of depth estimation.
DownLoad: