Cascaded CNN method for far object detection in outdoor surveillance

In maritime surveillance, detection of small ships and vessels located far away in the scene is of vital importance for behaviour analysis. Comparing to closely located objects, far objects are often captured in a smaller size and lack the adequate amount of details. Therefore, conventional detectors fail to recognize them. This paper proposes a CNN-based cascaded method for reliable detection of objects and more specifically vessels, located far away from a surveillance camera. The cascaded method improves small object detection accuracy by additional processing of the obtained candidate regions in their original resolution. The additional processing includes another detection iteration and a sequence of detection verification steps. Experimental results on our real-world vessel evaluation dataset reveal that the cascaded method increases the recall rate and F1- measurement by 13% and 12%, respectively. Another benefit is that the method does not require an adopter to change the model and architecture of the applied network. As an additional contribution, we provide a labeled maritime dataset to open public access.