Detection of Oversized Objects in a Video Stream Using an Image Classification with Deep Neural Networks

The applicability of an image classification approach to detect large anomalous and dangerous objects in a video stream was investigated for a safe opencast mining process. Using this method, an oversized rock detection system was developed on a conveyor belt in an opencast mine. Deep neural networks were successfully used to classify a frame containing a dangerous rock, e.g. ResNet, as well as VGG16. A detection algorithm based on a moving average thresholding was proposed in order to reduce false positives.