Artificial intelligence based anomaly detection for camera-based perception systems

Camera-based detection systems have become systems used in many critical areas and tasks today. Errors and anomalies that occur on the camera can cause major problems or damage for humans, products, and the environment. For the security and safety of camerabased systems, it is important to detect errors and anomalies. Within the scope of the thesis work titled " Artificial Intelligence based Anomaly Detection for Camera-based Perception Systems", the aim is to classify the noise types hat cause critical errors and anomalies in camera-based systems using artificial intelligence methods. In this context, the Automated robot control cell for quality control of automotive body-in-white, which was considered within the VALU3S project, was selected as a usage case for error and anomaly detection. This thesis work focuses on simulating errors and anomalies with noise from hardware or software on the images obtained from the robotic camera detection system by using the digital twin of a bus chassis in a simulation environment. During this process, a data set including different errors and anomalies was constructed by adding noise that is frequently encountered in real life to the images obtained through the bus chassis control system. In the first phase, classifications were made on this data set by using the Common Vector Approach (CVA), Convolutional Neural Networks (CNN), and CVA_based CNN models for four different noise classes. In the second phase, two new types of image noises were added to the data set and tests were conducted with the same classifiers for six different classes. The results obtained are presented in this study. The studies have shown that the CVA integrated CNN method provides better performance than when CVA and currently proven CNN methods such as Resnet, EfficientNet are used separately in terms of noise recognition, training and test times. The comparative tables with tests conducted for four and six different classes have been provided for noise detection success