Unusual event analysis for deep sea submersible

摘要

One of the main task for deep sea submersible is for event / object observation, e.g., new species fish or shrimp, strange topography. In this paper, by concerning deep sea animal motion or any interesting event as unusual event, we propose a new visual framework for unusual deep sea event analysis, which intends to reduce the onboard crew labor and improve the accuracy and efficiency accordingly. In comparison with most state-of-the-arts focus on fish tracking, ours contains much more diverse functions including unusual event detection, tracking and summarization. All these tasks are based on Chinese deep sea submersible, Jiaolong, mounted several video cameras around it. Specifically, for the PTZ camera, our framework can first automatically detect the unusual event by visual saliency, and track the corresponding object with the our previous online learning tracker; moreover, human operator can manually re-initialize it anytime to achieve human-in-loop control. For other stationary camera, we extract the key frames by our video summarization with group sparsity. To justify the efficiency and effectiveness of our proposed visual framework, a new deep sea unusual event dataset is collected from the offline recorded Jiaolong video cameras, and annotated by ourselves for a fair evaluation. Experimental results are reported based on our own dataset, where our method can detect, track and summarize the unusual event properly.