以行動深度學習網路偵測車輛之車道變換輔助系統

A Driver-Assistance System for Lane Change Based on Vehicle Detection with Mobile Deep Learning Networks

陳葦(Wei Chen)

指導教授 ： 繆紹綱

中原大學/電機資訊學院/通訊工程碩士學位學程/碩士(2018年)

https://doi.org/10.6840/cycu201800624

摘要

隨著科技的進步，許多車輛安全輔助系統被提出。大多數系統都是針對駕駛前方的視野進行威脅判定，針對後方來車的系統相對較少，當駕駛要變換車道時，除了注意與前方車輛距離的同時，也需要考慮後方車輛是否會對自身車輛造成威脅。基於上述原因，本研究提出利用車載攝影機所得之影像，用來偵測後方車輛威脅，藉此評估變換車道的可行性，提供建議給駕駛判斷行車狀況，減少駕駛人對於後方威脅的精神負擔，使駕駛人可以更專注於前方的威脅，提升行車的安全性。
本研究將輸入的行車影像辨識為白天或夜晚，區別出白天與夜晚後，分別針對這兩種模式進行左右車道擷取，同時消除複雜的背景環境，以免造成系統的誤判。接著需要偵測車輛的位置，若判斷為白天模式，則使用行動深度神經網路來進行車輛偵測；若判斷為夜晚模式，系統會將影像從原先的RGB色彩空間轉換成HSV色彩空間，並在HSV色彩空間進行車燈偵測，利用車燈找出車的位置。得到後方車輛的位置後，將該車輛與我方車輛間的像素距離轉換成實際距離。得到實際距離後，計算一定時間間隔車輛之距離變化，估算出後方車輛的相對速度。最後進行來自後方車輛的威脅判定，將相對速度帶入煞車距離公式與車輛在當下的實際距離進行比較後，給出駕駛一個當下是否變換車道的建議。
實驗結果顯示，不論是在白天或是夜晚都能夠順利偵測出車輛。距離與相對速度的估算，會因為曲線方程式造成些微的誤差，產生誤差的部分主要位於較遠、尚未造成威脅的車輛，因此在駕駛人的反應中是能夠被克服的。整體來說，系統能夠提供有效的意見，幫助駕駛在變換車道的同時減少負擔，提高行車安全。

關鍵字

深度學習網路 ； 車輛偵測 ； 距離估測 ； 曲線擬合 ； 變換車道

並列摘要

As technology evolves, many vehicle safety assistance systems have been introduced, and most of them put emphasis on the threat assessment in the front instead of the back. However, when a driver wants to change lanes, besides paying attention to keep a safety distance from the vehicle in front, the driver also needs to consider whether the vehicle behind will pose a threat to the driver’s vehicle. For these reasons, we proposed a method which used the images captured from a vehicle on-board camera to detect the vehicle threat from the back. With this technique we can assess the suitability of changing lanes and provide advice to a driver on making a lane change decision, thereby reducing the driver's mental burden for the threat from the back. The driver can focus on the threats in front of the own vehicle and improve the driving safety.
Our system will classify an image into the category of day or night. After differentiating day and night, the image covering left and right lanes is captured and the complex and irrelevant background is removed to avoid misjudgment of the system. Next, we need to detect the location of the vehicle. If day time is determined, the system will enter the day-time mode and use the mobile deep learning network to detect the vehicle. If night time is determined, the system will convert the original image in RGB color space to that in HSV color space to detect the head light of vehicles and find the location of the vehicle accordingly. Given the vehicle location, the system will transform an image pixel count, corresponding to the distance between the driver's car and the vehicle behind it, into actual physical distance. Then the system calculates the distance change in a fixed time interval to estimate the relative speed between the two vehicles involved. The vehicle threat from the back is assessed based on the current distance between the two vehicles involved, their relative speed and the brake stoppable distance. Finally, according to the level of the threat, we can provide the lane change suggestion for the driver.
Experimental results show that vehicles can be detected successfully both in day and night. Although the curve fitting function for pixel and actual distance transformation will cause a small error in distance estimation, it can be compensated by the driver’s response because larger errors mainly come from the situation of great distance and low threat. In the appropriateness of lane change, the system can provide an effective suggestion. Overall, the system can reduce the driver's mental burden by providing the effective suggestion on lane change decision and achieve the goal of driving safety.

並列關鍵字

deep learning network ； vehicle detection ； distance estimation ； curve fitting ； lane change

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。

參考文獻

[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。

[3] M. Oliveira and V. Santos, "Automatic Detection of Cars in Real Roads using Haar-like Features," in 8th Portuguese Conf. on Automatic Control, 2008.

[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, 2015