Fast obstacle detection system for the blind using depth image and machine learning

Abstract

Our research proposes a novel obstacle detection and navigation system for the blind using stereo cameras with machine learning techniques. The obstacle classification result will navigate users through a difference directional sound patterns via bone conductive stereo headphones. In the first stage, the Semi-Global block-matching technique was used to transform stereo images to depth image which can be used to identify the depth level of each image pixel. Next, fast 2D-based ground plane estimation which was separate obstacle image from the depth image with our Horizontal Depth Accumulative Information (H-DAI). The obstacle image will be then converted to our Vertical Depth Accumulative Information (V-DAI) which was extracted by a feature vector to train the obstacle model. Our dataset consists of 34,325 stereo-gray images in 7 different obstacle class. Our experiment compared various machine learning algorithms (ANN, SVM, Naïve Bayes, Decision Tree, k-NN and Deep Learning) performance between classification accuracy and prediction speed. The results show that using ANN with our H-DAI and V-DAI reaches 96.45% in obstacle classification accuracy and 23.76 images per second for processing time which is 6.75 times faster than the recently ground plane estimate technique.