Integrity Assessment of Maritime Object Detection Impacted by Partial Camera Obstruction.

The performance and usage of machine learning based object detection in visual data has increased significantly in the past decade. This technology can enable automated decision making in various applications, extracting key information in real time from a camera-based monitoring solution. However, to ensure the resilience and dependability in a safety-critical application, the intelligence provided by the monitoring solution must be reliable. In this context, understanding the impact of potential disturbances in the object detection performance is important. This work conducts an integrity assessment of maritime object detection under partial camera obstruction events. A subset of the ShipSG dataset, which contains thousands of annotated and segmented ship images from a harbor, and the Faster-RCNN object detection algorithm, trained on the seven different ship classes of the dataset, were used. The effect of simulated obstructions, of various intensities and configurations, on the false-positive, misclassification, false negative ratios, and associated detection score distributions were investigated. The outcome suggests that the use of a partial obstruction detection step, and the consideration of that information, can mitigate the consequences of faults in the object detection.