Vote3Deep: Fast object detection in 3D point clouds using efficient convolutional neural networks

Abstract

This paper proposes a computationally efficient approach to detecting objects natively in 3D point clouds using convolutional neural networks (CNNs). In particular, this is achieved by leveraging a feature-centric voting scheme to implement novel convolutional layers which explicitly exploit the sparsity encountered in the input. To this end, we examine the trade-off between accuracy and speed for different architectures and additionally propose to use an L1 penalty on the filter activations to further encourage sparsity in the intermediate representations. To the best of our knowledge, this is the first work to propose sparse convolutional layers and L1 regularisation for efficient large-scale processing of 3D data. We demonstrate the efficacy of our approach on the KITTI object detection benchmark and show that Vote3Deep models with as few as three layers outperform the previous state of the art in both laser and laser-vision based approaches by margins of up to 40% while remaining highly competitive in terms of processing time.

SARPNET: Shape attention regional proposal network for liDAR-based 3D object detection

“…Point cloud based 3D object detection. Voxel based methods [3,9,30,25] share a main idea to project sparse point cloud into compact representation. VoxelNet [35] employs VFE layers based on PointNet for 3D space points characterization.…”

Range Adaptation for 3D Object Detection in LiDAR

“…Object detection in point clouds is an intrinsically three dimensional problem. As such, it is natural to deploy a 3D convolutional network for detection, which is the paradigm of several early works [3,12]. While providing a straightforward architecture, these methods are slow; e.g.…”

“…While providing a straightforward architecture, these methods are slow; e.g. Engelcke et al [3] require 0.5s for inference on a single point cloud. Most recent methods improve the runtime by projecting the 3D point cloud either onto the ground plane [2,10] or the image plane [13].…”

Low Resolution Lidar-Based Multi-Object Tracking for Driving Applications

“…Deep learning techniques and Convolutional Neural Networks have been applied with great success to classical computer vision problems such as object classification [13] [11], detection [22], [23], and semantic segmentation [17]. Some early approaches on using CNNs to detect vehicles over 3D lidar point clouds make use of 3D convolutions [14] or sparse 3D convolutions acting as voting weights for predicting the detection scores [8,10]. However, due to the high dimensionality and sparsity of 3D lidar data, deploying them over point clouds implies high computational burden.…”

“…However, recent works [8], [5] are pointing at Deep Learning techniques as powerful tools to extract information from point clouds, expanding their applicability beyond image processing tasks. In previous works [26], we developed a vehicle lidar-based tracking system that used a Fully Convolutional Network (FCN) to perform per-point data segmentation using a Velodyne HDL-64 sensor.…”