基於深度學習之道路障礙物偵測與盲人行走輔助技術

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/84149

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/84149

题名:	基於深度學習之道路障礙物偵測與盲人行走輔助技術
作者:	沈鴻儒;Shen, Hong-Ru
贡献者:	電機工程學系
关键词:	穿戴式裝置;避障控制;深度學習;單眼深度估測;物件偵測;Wearable device;Obstacle avoidance control;Deep learning;Monocular depth estimation;Object detection
日期:	2020-07-21
上传时间:	2020-09-02 18:23:42 (UTC+8)
出版者:	國立中央大學
摘要:	本論文使用深度學習技術進行戶外障礙物辨識與障礙物距離偵測，同時設計一套戶外穿戴式導盲裝置，提供視障人士於戶外行走時更為安全、輕便的導盲系統。障礙物距離偵測分兩種方式，其一為使用單眼攝影機透過單眼深度估測神經網路輸出原影像之視差影像，依照不同視差區間使用不同迴歸方程式，分段將視差影像轉換為深度影像。障礙物偵測可使用語義分割神經網路或物件偵測神經網路，其中物件偵測神經網路經修改後，可多預測偵測框旋轉角度，使偵測框更加貼合障礙物。結合障礙物偵側結果，將單眼影像上障礙物偵測區域依據閥值篩選機制輸出障礙物距離數值，再利用障礙物距離數值計算出障礙物高度與方位。結合上述障礙物距離偵測結果，單眼深度搭配物件偵測應用於穿戴式導盲裝置，當影像裡出現物件偵測未能辨識之障礙物類型，將以單眼深度搭配語義分割非道路之區域偵測障礙物距離，進行後續避障控制；其二為使用雙眼攝影機計算出原影像之深度影像，使用物件偵測辨識障礙物，將影像上障礙物偵測區域中每一像素深度值由小排列至大，取其第一四分位數作為障礙物距離數值，雙眼深度搭配物件偵測同樣應用於穿戴式導盲裝置，可進行避障控制，同時運算速率高，可整合招牌追蹤系統。基於安全考量，視障人士行走於街上需靠著右側行走，本論文設計一套靠右行走演算法，依據相機擺放高度與相機參數，使用透視投影法找出實際深度與寬度投影至影像平面上之座標點，依此法畫出距離道路右側寬度之參考線，以及使用者左半身寬度之參考線，形成兩側寬度參考線，搭配語義分割之道路區域畫出道路邊線，依照兩側寬度參考線與道路邊線之間相對關係，提醒視障人士進行左右修正、直走、迴轉等動作。同時，投過前述之障礙物資訊，設計一套避障控制方法，依據障礙物位於前方之距離、高度與方位，執行避開障礙物、跨過障礙物或是停止等動作。綜合障礙物辨識與距離偵測，以及靠右邊走的控制，將可協助視障人士安全的走在馬路右側，避開障礙物，前往目的地。 ;In the thesis, deep learning technology is used for outdoor obstacle identification and obstacle distance detection. Meanwhile, a set of outdoor wearable guide device is designed to provide a safer and more portable guide system for visually impaired people during outdoor walking. Obstacle distance detection points in two ways, one is for the use of monocular camera through monocular depth estimation neural network to get the disparity image of the original image, through the regression analysis to the disparity image is converted to a depth image, image on obstacle detection area through the histogram statistics way to calculate obstacle distance output value, then calculate the obstacle height and position according to the obstacle distance. Obstacle detection uses semantic segmentation neural network or object detection neural network, calculate the obstacle distance from the obstacle identification results. Among them, after modification of the neural network of object detection, the rotation angle of the bounding box can be predicted to make the bounding box fit the obstacle better. Combined with the above obstacle distance detection results, the semantic segmentation collocation monocular depth is applied to the blind guide robot, and the object detection collocation monocular depth is applied to the wearable guide device for subsequent obstacle avoidance control. Second for the use of the stereo camera to calculate the depth image of the original image, using the object detection to identify obstacles, image on obstacle detection area in depth of each pixel values arranged from low to high, take the first quartile as obstacle distance, depth from stereo camera collocation object detection is also used in wearable guide device for the obstacle avoidance control, but high speed of operation, at the same time can be integrated signboard tracking system. Based on security considerations, visually impaired people need to walk on the right side of the road. The thesis designs a keep to right side algorithms, according to the camera position height and the camera intrinsic, use the perspective projection method to find out the actual depth and width of the projection to the coordinate of image plane. According to this method, drawing the reference line of the width distance from road to user, as well as the user′s left half body width of the reference line, formed the reference line of both sides. Match the semantic segmentation to locate the road area , in accordance with the relative relationship between width on both sides of the reference line and road edge, reminding the visually impaired that do corrections by going straight, moving to the left or right, and the rotation. At the same time, based on the obstacle messages, setting of obstacle avoidance control method is designed to perform actions such as avoiding obstacles, stepping over obstacles or stopping walking, according to the height, orientation and distance from the obstacles. Combined with each algorithm, leading the visually impaired to the destination.
显示于类别:	[電機工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	167	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....