A Review of Image and Point Cloud Fusion in Autonomous Driving

In the task of autonomous driving perception scenarios, multi-sensor fusion is gradually becoming the current mainstream trend. At this stage, researchers use multimodal fusion to leverage information and ultimately improve target detection efficiency. Most of the current research focus on the fusion of camera and LIDAR. In this paper, we summarize the multimodal-based approaches for autonomous driving perception tasks in deep learning within the last five years. And we provide a detailed analysis of several papers on target detection tasks using LiDAR and cameras. Unlike the traditional way of classifying fusion models, this paper classifies them into three types of structures: data fusion, feature fusion, and result fusion by the different stages of feature fusion in the model. Finally, it is proposed that the future should clarify the evaluation index, improve the data enhancement methods in different modes, and use multiple fusion methods in parallel in the future.