Abstract
A reliable obstacle detection system is crucial for autonomous surface vehicles (ASVs) to realize fully autonomous navigation with no need of human intervention. However, the current detection methods have particular drawbacks, such as poor detection for small objects, low estimation accuracy caused by water surface reflection, and a high rate of false-positive on water-sky interference. Therefore, we propose a new encoder-decoder structured deep semantic segmentation network, which is a water obstacle detection network based on image segmentation (WODIS), to solve the abovementioned problems. The first design feature of WODIS utilizes the use of an encoder network to extract high-level data based on different sampling rates. In order to improve obstacle detection at sea-sky line areas, an attention refine module (ARM) activated by both global average pooling and max pooling to capture high-level information has been designed and integrated into WODIS. In addition, a feature fusion module (FFM) is introduced to help concatenate the multidimensional high-level features in the decoder network. The WODIS is tested and cross-validated using four different types of maritime datasets with the results demonstrating that the mean intersection over union (mIoU) of WODIS can achieve superior segmentation effects for sea-level obstacles to values as high as 91.3%.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have