Title:	Curb Detection for a Pedestrian Assistance System using End-to-End Learning
Authors:	Qureshi, Hasham Shahid Wizcorek, Rebecca
Citation:	Journal of WSCG. 2018, vol. 26, no. 1, p. 75-81.
Issue Date:	2019
Publisher:	Václav Skala - UNION Agency
Document type:	článek article
URI:	http://hdl.handle.net/11025/35589
ISSN:	1213-6964 (on-line) 1213-6972 (print) 1213-6980 (CD-ROM)
Keywords:	hluboké učení;detekce obrubníků;asistenční systém pro chodce;učení se od začátku do konce;Leddar;monokamera;sloučení dat více senzorů;konvoluční neuronové sítě;umělé neuronové sítě
Keywords in different language:	deep learning;curb detection;pedestrian assistance system;end-to-end learning;Leddar;monocamera;multi sensor data fusion;convolutional neural networks;artificial neural networks
Abstract in different language:	Our goal is to develop an assistance system for supporting road crossing among older pedestrians. In order to accomplish this, we propose detecting the curb stone from the pedestrians’ point of view. Curb detection plays a significant role in road detection and obstacle avoidance, etc. However, it also presents significant challenges such as the small size of the target as well as, obstacles and different structures. To tackle these problems, we chose to fuse two sensors, a Camera and a Leddar, and use an algorithm that applies an end-to-end learning approach. The convolutional neural network was chosen to process the images acquired from the mono camera by filming the curb and its surroundings. The artificial neural network was selected to process the point cloud data of the Leddar acquired in the form of arrays from the 16 channels of the Leddar. A prototype was developed for data collection and testing purposes. It consists of a structure carrying both sensors mounted on a walker. The data from both sensors were collected with multiple factors taken into consideration, such as, weather, light conditions and, approaching angles. For the training of algorithms, an end-to-end learning approach was selected where we labelled the complete image or array rather than labelling the individual pixels or points in the data. The networks were trained and, the features from the parallel networks were concatenated and given as the input to the fully connected layers to train the complete network. The experimental results show an accuracy of more than 99% and robustness of the end-to-end learning approach. Both sensors are relatively inexpensive and are in fusion together, they are able to efficiently accomplish the task of detecting the curb stone from the pedestrians’ point of view.
Rights:	© Václav Skala - UNION Agency
Appears in Collections:	Volume 27, Number 1 (2019)

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.