3D Object Detection Based on LiDAR Data

Object detection has been a very hot research topic since the advent of artificial intelligence and machine learning. Its importance is very high specifically in advancing autonomous vehicles technology. Many object detection methods have been developed based on different types of data including image, radar, and lidar. Some recent works use point clouds for 3D object detection. One of the recently presented efficient methods is PointPillars, an encoder which learns from data in a point cloud and organizes a representation in vertical columns (pillars) for 3D object detection. in this work, we use PointPillars with lidar data of some urban scenes provided in nuScenes dataset to predict 3D boxes for three different classes of objects (car, pedestrian, bus). We also use nuScenes detection score (NDS) which is a consolidated metric for detection task, to measure and compare different scenarios. Results show that by increasing the number of lidar sweeps, the performance of the 3D object detector improves significantly. We try to increase the performance of the encoder by developing a method to combine different types of input data (lidar, radar, image) based on a weighting system and use it as the input of the encoder.