Obstacle Detection by Applying Deep Learning for Assisting the Driver of Mobile Scooter

Abstract

In general, unsupervised learning is essential for the anomaly detection problem because it is difficult to collect anomaly data. As for anomaly detection algorithms using images, the SPADE algorithm, which has been tried on various industrial products, is used to detect anomalies using k-NN with only normal data as a prototype. The algorithm is applied to road surface images, where the concept of normality and abnormality is quite different from that of industrial products. The authors have been conducting research on image-based safety systems for mobility scooters, which are mainly used by elderly people and people with disabilities. In this paper, we experimentally confirm whether the system can detect objects with indefinite shapes such as garbage nets, plastic bottles, stands with signs or chains, and pole wires on the road, which cannot be learned by supervised learning due to the uncertainty of shapes.