UNMC-VIER AutoVision Database King Hann LIM, Anh Cat LE NGO, Kah Phooi SENG, Li-Minn ANG Faculty of Engineering The University of Nottingham, Malaysia Campus Jalan Broga, 43500 Semenyih, Selangor Darul Ehsan. {keyx7khl, kezklma, kezkps}@nottingham.edu.my Abstract — Designing a driver assistance system has become a trend in the automotive technology to improve the security and efficiency of driving. However, there is no standard on- road database to verify the performance and effectiveness of algorithms. In this paper, an automotive vision database is created to assist researchers analyzing their designed algorithm in a more convincing way. The UNMC-VIER AutoVision database composes of a series of single view videos embodying the information of traffic signs, vehicles and single/multiple lanes. In addition, multi-views videos, with the aid of three cameras are included in the database providing more visual information in the panoramic view of traffic scene for analysis. The standard setup and calibration of the database is discussed in the paper. Some applications are discussed along with the use of the database. Keywords Automotive; database; traffic sign; lane detection; vehicle detection. I. INTRODUCTION Driver Assistance Systems (DASs), with the aid of optical sensors have recently gained much attention in the automotive technology in order to achieve autonomous driving. These systems which involve partial human interaction, are able to automatically sense nearby conditions (eg. vehicle overtaking, the presence of traffic signs etc.), understand physical road environment and immediately trigger a warning system to alert a driver on possible collision, traffic rules violation, and other worse case scenarios [1]. Although these systems have been under intensive research for last two decades, the performance of DASs is still degrading sharply when it encounters unexpected road circumstances. Unlike to the most of the intelligent visual systems, DASs are mainly operated under various real-time outdoor environments. The real-time scenes have posed a lot of challenges such as rapid change of illumination, variability of weather, existence of arbitrary objects on roads etc. Moreover, the performance for individual algorithm is not comparable to other algorithms due to the disparity of the automotive database. This may lead to unconvincing situation of creating a new algorithm for DASs. The Vision and Autonomous System Center’s Image (VASC) Database, created by Carnegie Mellon University (CMU) [2] obtains some road image sequences taken from the Navlab series of vehicles. It involves several variable factors such as day or night road scenes, sunny or rainy weathers, and shadowy lanes. They mainly develop the database for the lane detection and road following algorithm. Some car samples [3] are provided by CMU for evaluating car detection algorithms. Additional real-time road pictures can be obtained from AARoads [4]. Besides, there is an existing traffic signs database [5] containing 48 images of size 360×270 in PNG format. It is designed to test for three classes, i.e. pedestrian crossing, compulsory for bikes and intersection sign where each class contains 16 images. More traffic signs are available in [6] for standard US highway signs. However, none of them completely presents the road environment. Therefore, a standard road database is important for evaluating a new designed algorithm compared with the existing DAS algorithms. To ease the process of comparing results, a real-time automotive vision database has been created in this paper. The Visual Information Engineering Research (VIER) group from the University of Nottingham Malaysia Campus (UNMC) has created an AutoVision database comprising of a series of traffic scene videos. The content of the single view videos embodies the information of traffic sign, vehicles and single/multiple lanes. At the same time, multi-view of traffic scene, with the fusion of three optical sensors, provides more on-road visual information for researchers to further investigate the road conditions. In the video content, the database contains variability of weather, lane conditions and various road sceneries. This database will help developing the DASs for algorithm evaluation and performance improvement. In general, the UNMC-VIER AutoVision database has following advantages: low-cost, extendable, comprehensive traffic scenes containing vehicles, traffic signs and lanes in single/multi-view, and video-based samples. This paper is organized as: Section 2 demonstrates some prerequisite equipment to create an automotive vision database. It is followed by the video camera calibration techniques of multi-view database creation. Section 4 describes the content of UNMC-VIER AutoVision database. Some examples are presented along with the use of the database in section 6 and followed by conclusion and future works. II. PREREQUISITE EQUIPMENTS The importance of having a real-time road database has motivated the creation of database in this paper. To capture the real-time traffic scene on road, a vehicle is a necessary for the database creation. A hatchback vehicle – Perodua Kelisa with 5 doors 989 cubic capacities (c.c.) is occupied during the whole database acquisition as can be seen in Fig. 1(a). The height of the car is 142 cm apart from the ground to the rooftop. The width of the car is 149 cm and the length is 349cm [13]. During database recording, three camcorders are utilized in order to generate a panoramic view of traffic scene. Three camcorders consist of one Panasonic SDR- S7 camcorder and two Panasonic SDR-SW20 camcorders. 2010 International Conference on Computer Applications and Industrial Electronics (ICCAIE 2010), December 5-7, 2010, Kuala Lumpur, Malaysia 978-1-4244-9055-4/10/$26.00 ©2010 IEEE 650