在各種天候下基於深度學習的車道變換駕駛輔助系統

在臺灣平均每三個人便擁有一輛車，而平時因出遊或工作等緣故，國道的使用率也很高。不過隱藏在便捷交通底下的，是因車速較快而導致交通事故的高死亡率，且根據國道公路警察局的統計，「變換車道或方向不當」一直是主要肇事原因。隨著時代進步，許多智慧駕駛輔助系統問世，但這類系統多是針對駕駛前方的安全，像是車道偏離警示系統和前車防撞警示系統等，對於變換車道後方警示的系統則較少。近年已有論文提出以行動深度學習網路偵測車輛之車道變換輔助系統，該系統根據車載後鏡頭影像偵測後方車輛，並判斷變換車道時其威脅程度，警示駕駛以提升行車安全。雖然該系統已有了一定的成果，不過在系統執行速度以及面對不良天氣狀況下的探討則顯得較為不足。臺灣除了時常下雨外，交通部高速公路局也公告了全台國道在每年冬季容易起霧的 21 處路段。由於我們所使用的深度學習網路，需要較為清楚的車輛輪廓才能有較高的辨識率，而不良天氣下取得的影像會有許多汙染(霧霾、雨滴)，造成系統效能下降，因此考量系統的完備性，勢必要克服這兩種天氣的狀況。本研究提出改善該系統之方法。在系統執行速度方面，將原先判斷日夜所需的影像處理改成直接擷取曙暮光資訊以減少計算量，並且比較不同的行動深度學習網路，找出最適合本系統的模型架構。另外針對臺灣較常遇到的兩種不良天氣，我們新增了雨天模式及霧天模式，在進行車輛偵測時，先將汙染影像進行影像復原，以此提高系統在天氣不佳下的準確度。實驗結果顯示，本研究所提系統能在提升準確度的同時達到即時的效果，且在雨霧天都能有效地提升準確度。總體來說，此系統能在危險接近時適時給出警告，以防駕駛一不留神的疏忽。

關鍵字

深度學習網路；距離估計；圖像修復；變換車道；車輛偵測

並列摘要

On average, every three people in Taiwan have a car, and the national highway is also used frequently because of travel or work. However, the convenient traffic is accompanied by the high death rate caused by traffic accidents due to the fast speed. According to the statistics from the National Highway Police Bureau, “improper lane or direction changing” has been the main cause of accidents. With the advancement of the times, many smart driving assistance systems have come out, but such systems are mostly aimed at the safety of driving ahead, such as the lane departure warning system and the front vehicle collision warning system, and there are fewer systems for warning from behind when changing lanes. In recent years, some study has proposed a lane change assisting system, where a mobile deep learning network was used to detect vehicles. That system detects the rear vehicle based on the rear-lens image of the vehicle, judges the degree of threat when changing lanes, and warns the driver if necessary, to improve driving safety. Although the system has achieved certain results, the corresponding study lacks of the discussion on execution speed of the system and the situation of facing bad weather. In addition to frequent rains in Taiwan, the Highway Bureau of the Ministry of Communications also announced 21 sections of the national highway that are prone to fogging each winter. Due to the deep learning network we use, we need a clearer vehicle profile to have a higher recognition rate, and the images obtained in bad weather will cause a lot of pollution (smog, raindrops), resulting in system performance degradation. Thus, considering the completeness of the system, it is necessary to overcome these two kinds of weather condition. This study proposes ways to improve the system. In terms of system execution speed, the original image processing required to determine day and night is changed to directly capturing the twilight information with reduced amount of calculation, and the different mobile deep learning networks are compared to find the model architecture most suitable for the system. In addition, we have added a rainy day mode and a foggy day mode for the two kinds of bad weather that we often encounter in Taiwan. When performing vehicle detection, we first perform image restoration on the contaminated image to improve the recognition accuracy of the system under poor weather. The experimental results show that the proposed system can achieve immediate results while improving accuracy, and can effectively improve accuracy in rainy and foggy days. Overall, the system can give a warning when the danger is close, in case the driver is inadvertently negligent.

並列關鍵字

Deep learning network ； Distance estimation ； Inpainting ； Lane change ； Vehicle detection

參考文獻

[1]林建男，一個用視訊處理技術判斷來自後方車輛威脅的車道變換安全輔助系統，中原大學通訊碩士學位學程碩士論文，2016。

Google Scholar

[2]藍昱淳，處理車載攝影機所得影像以評估來自後方車輛的安全威脅，中原大學通訊碩士學位學程碩士論文，2017。

Google Scholar

[3]陳葦，以行動深度學習網路偵測車輛之車道變換輔助系統，中原大學通訊碩士學位學程碩士論文，2018。

Google Scholar

[4]TY. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, and J. Hays, "Microsoft COCO：Common Objects in Context," arXiv preprint arXiv:1405.0312, Feb. 2015.

Google Scholar

[5]N. Dalal, B. Triggs "Histograms of Oriented Gradients for Human Detection," in 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), Jun. 2005.

Google Scholar

國際替代計量

在各種天候下基於深度學習的車道變換駕駛輔助系統

未授權

主題瀏覽