Obstacle Classification Based on Laser Scanner for Intelligent Vehicle Systems

Obstacle Classification Based on Laser Scanner for Intelligent Vehicle Systems

Danilo Caceres Hernandez (Universidad Tecnológica de Panamá, Panama), Laksono Kurnianggoro (University of Ulsan, South Korea), Alexander Filonenko (ABBYY, Russia) and Kang-Hyun Jo (University of Ulsan, South Korea)
DOI: 10.4018/978-1-5225-9924-1.ch010

Abstract

In the field of advanced driver-assistance and autonomous vehicle systems, understanding the surrounding vehicles plays a vital role to ensure a robust and safe navigation. To solve detection and classification problem, an obstacle classification strategy based on laser sensor is presented. Objects are classified according the geometry, distance range, reflectance, and disorder of each of the detected object. In order to define the best number of features that allows the algorithm to classify these objects, a feature analysis is performed. To do this, the set of features were divided into four groups based on the characteristic, distance, reflectance, and the entropy of the object. Finally, the classification task is performed using the support vector machines (SVM) and adaptive boosting (AdaBoost) algorithms. The evaluation indicates that the method proposes a feasible solution for intelligent vehicle applications, achieving a detection rate of 87.96% at 48.32 ms for the SVM and 98.19% at 79.18ms for the AdaBoost.
Chapter Preview
Top

Introduction

A combination of environmental issues and demand for safety factors influence the automotive industry. The automotive industry has currently shown robust growth which has been reflected in the increase of access to data and information resulting in development of sensor solutions. Advance driver assistance systems (ADAS) and autonomous ground vehicle navigation (AGVN) are still facing important challenges in the field of robotics and automation. Essentially due to the uncertain nature of the environments, moving obstacles, and sensor fusion accuracy. In that sense, the systems must be able to recognize features in which the method should be able to detected lanes, objects on the road, e.g vehicle, pedestrian or animal (Zhang, F. et al.2016; Ibarra-Arenado et al. 2017; Wang, T et al. 2003). Zhang, F et al. proposed the use of LiDAR to develop a vehicle detection method. The authors presented a probability hypothesis density filter which is a multiple-target filter and for the case of hypothesis verification the authors used SVM. To evaluate their proposed idea the authors used the KITTI dataset (Geiger, A et al. 2013). Although the result shows a good performance, there are still issues with processing time due to the 3D Velodyne point clouds. Authors in (Ibarra-Arenado et al. 2017) proposed a vision-based vehicle detection method, which is comprised of two steps: hypotheses generation (HG) and hypotheses verification (HV). The authors are focuses on the HG strategy. In HG vehicles are localized by using a shadow-based vehicle detection method, distinguishing the difference on intensity between the shadow under the vehicle and road surface. As a results this method overcome the problems given by lateral shadows, asphalt stains and traffic markings on the road. However, the method still encounters problems of false positives rate mostly due to the ego-vehicle trajectory. Wang, T. et al. (2003) presented a radar-vision fusion based strategy. The idea focuses on the radar target detection of candidate objects. Once the candidate is detected, the coordinate information is used to define the searching are within the image where the object is located. The idea is to design a robust object detection; however, in order to be used in open road scenarios the ongoing idea need to be improved. Current research efforts in the field of obstacle detection and recognition strive to understand and improve the problem of features extraction stage by leaping between different sensors and strategies. Regarding the above paragraph, the most utilized sensors for achieving ADAS and AGVN using feature-based strategy (color, edges, shape, geometric) are commonly cameras and laser.

Complete Chapter List

Search this Book:
Reset