Object Detection on Video for Shipborne Environmental Perception System
Wusheng Shang, Jie Yang, Qingnian Zhang, Zhiqiang Guo · 2019
We present an application based on convolutional neural networks (CNN) as the first step of environmental perception, to identify objects nearby unmanned vessels. A Single Shot MultiBox Detector (SSD) architecture was built, with a MobileNet network as feature extractor. To test the feasibility, three datasets were generated and applied to our CNN model as training data. The trained model was used to identify ships in videos. Videos were cut into frames, and applied to the trained model. After objects in each frame were recognized, new videos with bounding boxes were generated as outputs. With the first two datasets, ships are successfully detected. Using the third dataset, a big cargo ship is well identified, while small ships are ignored for lack of training data. Results indicate that with a sufficient training, our implementation has good object detection performance, and verified to be suitable for an environmental perception system.