Object Detection and Classification in Occupancy Grid Maps Using Deep Convolutional Networks

Abstract

Detailed environment perception is a crucial component of automated vehicles. However, to deal with the amount of perceived information, we also require segmentation strategies. Based on a grid map environment representation, well-suited for sensor fusion, free-space estimation and machine learning, we detect and classify objects using deep convolutional neural networks. As input for our networks we use a multi-layer grid map efficiently encoding 3D range sensor information. The inference output consists of a list of rotated bounding boxes with associated semantic classes. We conduct extensive ablation studies, highlight important design considerations when using grid maps and evaluate our models on the KITTI Bird's Eye View benchmark. Qualitative and quantitative benchmark results show that we achieve robust detection and state of the art accuracy solely using top-view grid maps from range sensor data.

Capturing Object Detection Uncertainty in Multi-Layer Grid Maps

“…Furthermore, we validate uncertainty estimation during the RPN and the output stage for different models and analyze their differences. We evaluate our work on the KITTI Bird's Eye View Evaluation benchmark allowing comparison to previously published work [8]. Finally, we reduce the number of box parameters and represent pose and shape uncertainty by a common shape based on collision probabilities.…”

An LSTM Network for Real-Time Odometry Estimation

“…The success of these applications are mainly due to the fact that a large amount of datasets has become available in the last years. In the context of intelligent vehicles, interesting work has been developed in several different fields: trajectory prediction [5] [6], mapping [7], control [10] and even end-to-end approaches [8] [9], where the car is controlled completely by a Deep Learning module.…”

Improving Map Re-localization with Deep ‘Movable’ Objects Segmentation on 3D LiDAR Point Clouds

“…The bird's eye (zenithal) view I BE is obtained over an area of 60 × 50 meters in front of the LiDAR sensor after carefully observing that roughly 95% of the annotated vehicles in Kitti are within these margins. Inspired by [5], [16], we generate a 2D grid with a resolution of 0.1 meters and project the cropped point cloud on it. We consequently obtain a bird's eye view I BE ∈ R H ×W ×C , where H = 600, W = 500, and C = 6, accounting for six different features: 1) a binary occupancy term with zero value if no points are projected in the cell and one otherwise; 2) an absolute occupancy term, counting the total number of points in the cell; 3) the mean reflectivity value of the points on the cell; and 4, 5, and 6) the mean, minimum and maximum height values of the points projected on the cell.…”

“…Similarly, [15] creates a front view projection of the polar LiDAR coordinates along with the reflectivity of each point, to segment vehicles by predicting the vehicleness confidence of each point. On the other hand, BirdNet [5], TopNet [16] or RT3D [17] make use of a bird's eye view projection of the point cloud, encoding different features on each cell.…”