Training and Performance Analysis of Vehicle Detection Neural Networks to Field Test and Simulation Datasets of Multi-channel Lidar

Abstract

In this paper, the training method used in a lidar-based, object detection algorithm is applied to different types of datasets, i.e., experimental driving data and virtual simulation data. Then, their performances are compared with respect to different key performance indexes(KPIs). Among many object detection methods introduced in the literature, three distinguished networks that consider the representation of lidar cloud points are chosen to compare fine tuning and performance. While most open datasets reflect only safe driving situations, it is necessary to develop and validate the object detection algorithm in dangerous and critical situations. With the generation of a virtual simulation dataset, including unsafe scenarios, the performance of the object detection algorithms can improve when the fine-tuning method is applied, along with the virtual dataset.