PillarFusion: Efficient 3D Fusion Object Detection Based on Pillars

With the rise of autonomous driving, the technology of object detection has developed rapidly. However, existing 3D object detection methods have some shortcomings. The detection accuracy of image-based methods is low, while point cloud-based methods have difficulty in overcoming the problem of data sparseness. Some recent multi-modal methods have high training costs and low inference speed, leading to difficulties in deployment. In this paper, we propose PillarFusion, an efficient multi-modal object detection method based on pillars. We first convert the 3D point cloud data into pseudo-images and pillars. Moreover, we extract the features of the image by using an image feature network. Then we fuse image and point cloud features and use a novel backbone network with 2D CNN to improve the detection speed. Finally, we output the predicted 3D bounding boxes. We validate our proposed method on the KITTI data set. The results demonstrate a significant improvement in average precision of Intersection over Union (AP@IoU) by 3.80% compared to existing methods. Meanwhile, our method improves the inference speed by at least 2.9 times.