We address the problem of real-time floating obstacle detection in aquatic environments with convolutional neural networks. Reliable and fast detection of obstacles is crucial for autonomous driving. Convolutional neural networks are often used in autonomous cars but have not yet been thoroughly tested in the aquatic environment. For this purpose, we analyze two of the latest convolutional neural networks for object detection and classification: YOLO and BlitzNet. We propose a modified convolutional neural network for obstacle detection YoloW and a new dataset for object detection in the aquatic environment. The dataset contains 19691 annotated obstacles appearing in 12168 images. We propose a customized learning process from uncertain training examples, which is suitable for training convolutional neural networks on real world datasets. We evaluate the performance of the presented convolutional neural networks on the proposed dataset. By increasing the number of training examples, the accuracy of all models is improved. After training on the entire training set of our dataset, BlitzNet achieves an average accuracy of 89.68%, YOLO 96.78%, while our model YoloW achieves an average accuracy of 97.72%. The proposed YoloW works in real-time and is capable of obstacle detection at 30.12 images per second on average.
|