Image-based anomaly detection using β -Variational Autoencoder for surface vehicle collision avoidance

Abstract: Unmanned vehicles need robust systems to ensure the safety of the vehicle and its environment. Being able to find and avoid perilous situations is paramount to such a system. In this paper we suggest an unsupervised image-based anomaly detection algorithm using a variational autoencoder and a superpixel segmentation algorithm, which is adapted to the maritime obstacle detection task. The algorithm locates potentially hazardous objects and calculates the distance to them by measuring the images’ reconstruction error over segmented regions. The algorithm’s results on the public MODD2 dataset shows that it has difficulties finding small objects and that it cannot compete with the current state-of-the-art supervised segmentation algorithms on the same dataset, with an F1 score of 26.6% compared to 82.7%. Although further research and optimization is required to utilize the algorithm in a production level product, the results indicate that the algorithm is worth investigating further as it is able to detect many of the objects in our testing videos and due to it having applications in several areas.