本論文主要基於電腦視覺技術的研究並且實作三種類型應用，包含人體姿態辨識、機器手臂的物件夾取與移動式機器人的地形辨識。在人體姿態辨識的部分，使用了Kinect深度影像感測器擷取影像，透過影像相減、形態學與物件聯通法等影像裡技術，進行前處理切割出人體的影像。再以人體影像的重心點位置，區分人體的上半身與下半身，藉由上、下半身最大寬度的比例，辨識當前姿態是否為跪姿。接著，使用星狀骨骼化擷取人體的特徵向量，配合Learning Vector Quantization (LVQ)類神經網路分辨當前姿態是坐姿、彎腰、躺姿、站姿或側坐。最後，因側坐與站姿的特徵向量非常相似，不容易被LVQ類神經網路分辨出，因此需要再比對其人體的寬度、高度的比例確認當前的姿態為站姿或是側坐。在機器手臂物件夾取的研究中，由於機器手臂結構與背隙的問題，難以準確地移動夾爪至目標物的位置。因此，採用雙眼視覺裝置辨識夾爪在空間中的實際位置，並且設計一個模糊控制器補償機器手臂位置的誤差，以提升機器手臂控制的精度。藉由本論文提出的補償方法，機器手臂可以準確的移動到目標物的位置，並且成功的夾取該物件。第三種研究使用了XtionPro深度影像感測器，擷取機器人前方地面的影像。接著，透過影像幾何的分析建立一個虛擬的平坦地形影像，做為機器人前方地形平坦度評估的標準。將建立的虛擬影像與實際擷取到的地面影像進行比對，比對的結果中差異性較大的像素點部分則歸類為不平坦的區域。藉由此方法可以利用感測器偵測出前方的障礙物與凹洞的位置，配合本研究提出的機器人動作策略，驅動機器人安全地到達目標點位置。;This dissertation presents three studies based on computer vision containing posture recognition, object grasping by a robot arm, terrain traversability estimation for wheeled mobile robot. In the first study, an effectivity posture recognition method is proposed based on depth image captured by Kinect sensor. Several image processing techniques are applied to extract the features on the human postures such as ratio of body and star skeleton. The ratio of upper and lower body width can fast distinguish the posture whether posture is kneeling or not. Then, Learning Vector Quantization neural network is used to recognize the four categories of human postures forward sitting, stooping, lying and the other. One more check is the final step of the posture recognition method to judge standing and non-forward sitting. The second study utilizes the stereo vision to enhance the accuracy of the robot arm without any external sensors. Due to the backlash, the gripper cannot approach the target object. The stereo vision is applied to recognize the actual position of the gripper and then the fuzzy control is adopted to compensate the position error. After compensation the robot arm can successfully grasp the target object demonstrated in the experiments. The third study develops and implements a fast terrain traversability estimation method using a depth image sensor XtionPro. In this study, a virtual terrain surface image is created and compared with captured upcoming terrain image to extract the features of the terrain. Based on the features, any obstacle and hollow are found. Then, a movement strategy is proposed for robot to make reaction to the obstacle and hollow and approach the goal position.
