Anomaly detection as vision-based obstacle detection for vehicle automation in industrial environment

Abstract

Nowadays, produced cars are equipped with mechatronical actuators as well as with a wide range of sensors in order to realize driver assistance functions. These components could enable cars’ automation at low speeds on company premises, although autonomous driving in public traffic is still facing technical and legal challenges. For automating vehicles in an industrial environment a reliable obstacle detection system is required. State-of-the-art solution for protective devices in Automated Guided Vehicles is the distance measuring laser scanner. Since laser scanners are not basic equipment of today’s cars in contrast to monocameras mounted behind the windscreen, we develop a computer vision algorithm that is able to detect obstacles in camera images reliably. Therefore, we make use of our well-known operational design domain by teaching an anomaly detection how the vehicle path should look like. The result is an anomaly detection algorithm that consists of a pre-trained feature extractor and a shallow classifier, modelling the probability of occurrence. We record a data set of a real industrial environment and show a robust classifier after training the algorithm with images of only one run. The performance as an obstacle detection is on par with a semantic segmentation, but requires a fraction of the training data and no labeling.