Blind Navigation in Outdoor Environments: Head and Torso Level Thin-Structure Based Obstacle Detection

Abstract

Blind navigation in computer vision is a highly active research area because independent mobility is one of the essential needs of every human being. Among many blind navigation and obstacle detection systems, researchers have given limited attention to the detection of thin structured wires like obstacles even though blind people can be severely damaged by them. In this research domain, a thinstructured wire is defined as a wire with a maximum of 5mm diameter. For finding a solution for this growing problem of detecting thin-structured wires for blind navigation in outdoor spaces, an assistive system was implemented. The main objective of this research is to address this research gap using computer vision-based techniques. The proposed approach is based on the Simultaneous Localization and Mapping algorithm(SLAM). SLAM is the computational problem of constructing or updating a map of an unknown environment while simultaneously keeping track of an agent’s location within it.The suggested approach takes input as a video stream of the path of the user from a monocular camera and that video stream is processed by the system frame by frame. This process consists of three main stages named image information extraction stage, Tracking stage, and Mapping stage. In the information extraction stage Difference of Gaussian(DoG) based edge detector is used to extract image edges and the outputs of the edge extraction stage are called keylines. Based on these keylines, edge map is created for each frame. Then in the tracking stage, camera motion is tracked by fitting the previous edge map into the new edge map using a warping function. The outcome of this stage is a Special Euclidean group (SE(3)) transformation and it is used as an input to the mapping stage. In the mapping stage, each keyline in the new edge map is matched against the ones in the previous edge map in order to filter real obstacle edges from the noise edges. Since there are no suitable benchmark datasets for evaluating the method, new datasets were created which consists of wires as obstacles in different environments. The system was evaluated using the Intel core I 3 machine without Graphics Processing Unit(GPU) support. According to the evaluation results, a maximum of 15 frames per second(fps) rate has been achieved with 75% accuracy for wire detection. In conclusion, the proposed algorithmic approach can be considered as a lightweight and accurate solution for the addressed research gap.