Pedestrian Detection from 3D Geometry in High Density Point Clouds

Author: Ian Pau Riera Smolinska

Virtual Room: https://eu.bbcollab.com/guest/5c54c0bdc57f41978b092feb1be43b72

Date & time: 22/09/2021 – 9:30 h

Session name: Self-driving vehicles

Supervisors: Josep R. Casas, Santiago Royo

Abstract:

With the irruption of autonomous driving in recent years, pedestrian detection has gained momentum. Despite the maturity of 2D detection on RGB images, there has been a tendency to add 3D sensors, such as Light Detection and Ranging (LiDAR), to complement data in situations where the 2D approach fails due to environment conditions. 3D detection datasets are still under construction and not as mature as for 2D object detection, and most of the state-of-the-art architectures rely on a single dataset: Kitti. Besides, the high computational cost of point cloud processing, caused most of the approaches to either exploit the 2D detection and back-project it to the point cloud, or to reduce its size by means of grouping points into voxels. The aim of this project is to explore how the characteristics of the input data can affect the performance of 3D detection, using solely the geometric information. To that end, we exploit PointRCNN, a point-based architecture that performs object detection directly over the raw point cloud data. We have annotated a pedestrian-oriented dataset captured with a L3CAM sensor from Beamagine, that provides a high density point cloud. The sensor also provides a synchronous RGB capture that helps in the annotation process. In this project we compare the detection results obtained using PointRCNN on the Kitti and Beamagine datasets.

Committee:

– President: Francesc Moreno-Noguer(UPC)
– Secretary: Dimosthenis Karatzas(UAB)
– Vocal: Antonio López Peña(UAB)