| This thesis aims to build a curb and edge detection method for autonomous wheelchair navigation using Convolutional Neural Networks (CNNs). In contrary to a deterministic approach, CNNs use a statistical inference approach to the classification problem. For this, a comprehensive dataset of curb and edge examples is collected. The training and test dataset was collected around Idaho State Universityand the Measurement and Control Engineering Research Center. After the neural network is trained using the dataset, it is tested using images not utilizedfor training.The entiresystem consists of the electric wheelchair,aKinect 360and a laptop. The Kinect 360 is used to capture images, and the laptop runs the CNN algorithm to classify the collected pictures. The configuration of the CNN is such that a number of feature detectors and a number of hidden layer neurons are varied in order to test the accuracy of the classification algorithm. An optimum configuration for the available training and test set is realized.The final CNN shows almost 100 percent accuracy for images with distinct curbs and edges. It still shows reasonable accuracy for images with partially hidden curbs or gently sloping edges.Key Words:Obstacle Detection, Vision Processing, Autonomous Vehicles, Autonomous Wheelchair,Convolutional Neural Networks, Kinect 360, Curbs and Edges Detection, Deep Learning, Keras, OpenNI, OpenCV |