Vessel detection in video with dynamic maritime background

Automating the detection of non-cooperative vessels in surveillance video is challenging. First, the detection algorithm has to handle a large degree of appearance variation of vessels with respect to shape, size and viewing geometry. Second, a unique challenge in the maritime domain is the presence of sea clutter, which can cause a high number of false detections. While recent research in object detection has largely been focused on objects on the ground, we have developed a layered detection algorithm to address challenges in the maritime domain by fusing cues from (1) a discriminative detection algorithm that learns a vessel target model from hundreds of vessel images, and (2) a dynamic texture-based background model that adaptively learns the spatiotemporal dynamics of sea clutter. We present results on how each layer of the algorithms was individually optimized, and how their outputs were fused. Initial results were promising showing a significantly lower false alarm rate than when only the target model was applied. The proposed approach has applications in port, coastal and waterway surveillance.