Abstract:

With the acceleration of urbanization, the intelligent transportation has received more and more attention. Among them, the use of depth completion technology to extract the depth information of objects plays an important role in the realization of vehicle target tracking, distance calculation between targets and other tasks. However, multi-source depth completion data collected in practice often have correlation bias, resulting in knotty errors. In this regard, this study studies the depth completion technology based on multi-source data association fusion. The proposed technology enhances the depth map by calculating multi-channel confidence, and performs more accurate point-by-point correlation between the image and the millimeter-wave radar point cloud data. By designing a multi-scale attention fusion module, the adaptive fusion of multi-granularity associated data is realized to generate high-quality depth maps. In this study, a large number of experiments have been carried out in the public nuScenes data set. The experimental results show that the mean absolute error of our method is 1.142 m, which is lower than the 1.472 m of the existing benchmark method.

Key words: intelligent transportation, multi-source data fusion, depth completion, deep learning, attention mechanism, adaptive combination