Object detection using depth completion and camera-LiDAR fusion for autonomous driving

Autonomous vehicles are equipped with complimentary sensors to perceive the environment accurately. Deep learning models have proven to be the most effective approach for computer vision problems. Therefore, in autonomous driving, it is essential to design reliable networks to fuse data from different sensors. In this work, we develop a novel data fusion architecture using camera and LiDAR data for object detection in autonomous driving. Given the sparsity of LiDAR data, developing multi-modal fusion models is a challenging task. Our proposal integrates an efficient LiDAR sparse-to-dense completion network into the pipeline of object detection models, achieving a more robust performance at different times of the day. The Waymo Open Dataset has been used for the experimental study, which is the most diverse detection benchmark in terms of weather and lighting conditions. The depth completion network is trained with the KITTI depth dataset, and transfer learning is used to obtain dense maps on Waymo. With the enhanced LiDAR data and the camera images, we explore early and middle fusion approaches using popular object detection models. The proposed data fusion network provides a significant improvement compared to single-modal detection at all times of the day, and outperforms previous approaches that upsample depth maps with classical image processing algorithms. Our multi-modal and multi-source approach achieves a 1.5, 7.5, and 2.1 mean AP increase at day, night, and dawn/dusk, respectively, using four different object detection meta-architectures.