AARMOD: Adaptive Attention Guided Robust Multimodal 3-D Object Detection

In the field of autonomous driving, robust multimodal 3-D object detection is still a crucial challenge for accurate and efficient 3-D object detection, particularly for vehicles in the occlusion environment. In this article, we develop a late fusion method for combining camera and LiDAR data to perform 3-D object detection, which is referred to as AARMOD. Specifically, we fuse the results of 2-D object detection with 3-D object detection and then integrate these fused outputs into a single network for 3-D detection. We adopt the voxel-based sparse embedded convolutional detection (SECOND) as the 3-D object detector and you only look once version 8 (YOLOv8) as the 2-D object detector. In order to improve the detection performance in the occlusion environment, we introduce the four-head adaptive spatial feature fusion detection head into YOLOv8 to effectively filter out conflicting information across different scales. And utilize the minimum point distance based intersection over union (MPDIoU) loss function which considers bounding box overlap, center distance, and aspect ratio, thereby reducing localization errors for vehicle targets. Additionally, due to the inherent challenges in effectively leveraging both camera and LiDAR modalities, we design an adaptive and simple attention module (ASimAM) into the fusion network, which enables the network to focus on the most relevant features by accurately capturing important information from both modalities. Finally, we conduct extensive experiments on the KITTI dataset, and the results demonstrate the efficacy of our method for accurate 3-D object detection in occluded environments.