Deep learning-based anomaly detection for imaging in autonomous vehicles

Autonomous driving and self-driven vehicles have become among the most pursued research areas in recent years. Nowadays, various driving tasks can be solved by applying the newest machine learning techniques such as line tracking, traffic sign recognition, automated speed adjustment, and parking. However, difficult visual conditions and anomalies can cause problems in selected algorithms, which may occur unexcepted and failure operations in these cases. It is also expected not just very expensive to do such kinds of experiments, but these problematic conditions are also lead to dangerous traffic situations at the same time. We made an effort to put these kinds of studies into a cost-effective and safe model-scale environment. This paper introduces an anomaly detection method capable of recognizing abnormal and burnt-out objects in image scenes. Our proposed method is based on a fast neural network architecture using YOLO layers to detect regions. Our experiments demonstrate the capabilities and detection accuracy of the designed neural network, called anomalyNet, with the complete training and evaluation process. In the study, we work with publicly available datasets, but our model-sized track and DAVE (University of Debrecen Autonomous VehiclE) play an important role also.