Real-time moving target detection in infrared maritime scenarios

In maritime surveillance, real-time moving target detection is a crucial task. The purpose of our work was to test some moving target detection techniques inspired by the state-of-the-art on a specific dataset of interest. In this paper, we analyze the performance obtained by Frame Difference and Gaussian Mixture Model-based methods in static InfraRed video sequences. The dataset was collected under real operational conditions during a recent experimental activity. The frames in the dataset are characterized by heterogeneous backgrounds and different targets covering a wide range of sizes and speeds. To evaluate the performance, the ground truth has been manually labeled through direct observation in 4706 frames. All the examined techniques are used to estimate the background. They are preceded by a frame-based z-normalization. After the background estimation, the absolute difference with the normalized frame is computed to highlight both hot and cold targets. Once highlighted, the targets can be separated from the background with a threshold. Then, morphological operations both in time and space are executed to delete small and brief false alarms. Finally, blob analysis is computed to extract the Regions of Interest. The algorithms have been evaluated based on their Precision-Recall curves, and Mean Execution Times.