基於深度學習的斜坡偵測應用於智慧型輪椅上坡安全控制器

Abstract

在老年化人口社會中，隨著自動化的發展，電動輪椅的使用率逐年增加，然而，使用者對於斜坡角度誤判造成行駛速度掌控不當，產生多起翻覆或輪椅下滑的意外事件。因此，本論文首先使用工研院所研發包含電池的薄型動力輪模組(LEAP)降低電動輪椅重量和體積的負擔，再使用嵌入式平台-TX2進行上坡影像辨識模組與控制器運算。 本論文中以模組化設計上坡辨識及安全控制器，上坡路段預先辨識分類透過RGB-D攝影機抓取環境深度資訊，採用深度學習(Deep Learning)技術訓練斜坡辨識模組，讓斜坡辨識不再受限於有明顯邊界和四個角點的斜坡路段，所提出的斜坡模組為CNN-4架構，相較於AlexNet和GoogleNet每一張圖像的辨識時間提升11倍，讓控制器更符合機器人即時控制的需求；加入遴選決策系統及陀螺儀傾角輔助，提升5%模組辨識率，達到97.1%。控制器基於Q-learning和調適性網路模糊推論系統(Adaptive Network-based Fuzzy Inference System)，依據使用者點選的命令及斜坡辨識結果自動調整行駛速度，透過Q-learning了解環境即時回饋以微調速度，加入斜坡預先辨識的上坡安全控制器相較無斜坡辨識的行駛時間少了23%，行駛斜坡所花費的總能量也減少51.79%，除了提升輪椅行駛於斜坡路段的效率外，也避免上坡瞬間加速太過劇烈造成翻覆的問題，以及對角度和速度誤判而產生的危險；碰到無法行駛的路段時，也會自動調減速度，警示使用者，整體提升上坡路段的行駛安全。

In an aging society, the demand for electric wheelchairs is growing with the development of automation. However, many accidents have occurred when the wheelchair drives on the ramp because of the mis-judgement of the slope and speed. This paper applies Light Electronic Assistance Pal (LEAP) to reduce the weight of electric wheelchairs. The modular design of uphill controller and ramp detection functions allows using users to easily employ the functions he/she selects. This paper proposes a ramp detection function implemented using deep learning algorithm with the CNN-4 structure to analyze depth data. The identification time of each video frame is 11 times faster than AlexNet and GoogleNet. The uphill safety controller is designed as a Q-learning and Adaptive Network-based Fuzzy Inference System. The safe speed is automatically calculated according to the angle of slope classification and revised in real-time during the slope driving to prevent the user from moving towards the dangerous ramp or rolling back due to inadequate power. The accuracy of ramp detection is increased by 5% to 97.1% due to assistance from the voting system processing and the gyroscope output of the wheelchair. The experiment of the uphill controller with ramp classifica-tion takes 20 seconds to complete the slope driving which is 23% less than the controller without uphill detection. The energy consumption is also one half less than the experiment without uphill detection.