Ultra-low-complexity Technologies for Advanced Driving Assistance System (ADAS)

• Main Authors: Li-Hung Wang, 王理弘
• Other Authors: 吳崇賓
• Format: Others
• Language: en_US
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/2yg895

博士 === 國立中興大學 === 電機工程學系所 === 106 === In this thesis, the ultra-low-complexity advanced driving assistance system (ULCADAS) is proposed to improve driving safety with high processing speed and acceptable accuracy. The ULCADAS is composed of pre-processing, lane departure warning (LDW) system, low-complexity vehicle detection (LCVD) system, and collision avoidance (CA) system. In pre-processing, different ROIs are proposed to reduce the redundant computation. In the ROIs, the block-based angle block classification and mean luminance calculation are applied to extract the features of lane markings and vehicles. The LDW system is composed of linear block-based lane marking identification and lane departure identification. Based on the classified angle blocks and the feature of lane markings in near distance, the lane markings are identified using the formula of the linear function in linear block-based lane marking identification. According to the slope of the identified lane markings, the lane departure is identified. The experimental results show that the detection rate, warning rate, and false alarm rate are 95.48%, 95.04%, and 3.85%, respectively. In LCVD system, the shadows are extracted using the mean luminance and the hypotheses are select from them according to the widths in different distances. The hypothesis is partitioned into non-overlapping regions and the descriptive feature is proposed to describe the distribution of angle blocks in each region. Comparing the hypothesis with created matching models, the verifying score is obtained. If the verifying score is higher than the threshold, the hypothesis is verified as a vehicle. The experimental results show that the precision, recall, and accuracy are 95.3%, 95.1%, and 90.8%, respectively. In CA system, the distance between the detected vehicle and the onboard camera is estimated by the homothetic triangles. Combining the vehicles with estimated distances and the result of LDW algorithm, the potential collision is identified. The experimental results show that the precision, recall, and accuracy of the proposed AC system are 94.78%, 94.62%, and 93.98%, respectively. The overall processing time of the proposed ULCADAS is 8.63 ms/f, which is equivalent to 115.87 fps and it is 3.86 times faster than that of real-time processing (30 fps).