Research On Water Surface Obstacle Detectionmethod Based On Semantic Segmentation Model

With people’s exploration and research on coastal waters,unmanned surface vehicle(USV)has been widely used as an important tool in the development of water surface environment,and use cameras as sensors to obtain optical images and perceive the surrounding environment.In order to ensure that the USV can navigate autonomously and safely and avoid collisions and potentially dangerous situations,water surface obstacle detection technology has become an important research content in image processing on the USV platform.Due to the complex and changeable water surface environment,the USV is affected by waves and the uncertainty caused by its own motion,which brings many challenges to obstacle detection.Among various water surface obstacle detection methods,the method based on the semantic segmentation model can realize obstacle detection and obtain better detection results,but the current method cannot completely solve the various problems in the detection.Therefore,in response to practical problems such as water reflections and blurred horizons,this paper proposes a water surface segmentation method based on an improved semantic segmentation model,and combines the saliency estimation detection method to complete water surface obstacles detection under monocular and binocular systems.The main research contents are as follows:(1)Aiming at the false detection of the reflective area on the water surface and the missed detection caused by the incomplete segmentation of small obstacles,a method based on the removal of reflective areas and saliency estimation is proposed.This method uses image preprocessing,semantic segmentation,and saliency estimation method realizes obstacle detection.First of all,by removing the reflective area in the water surface in the pre-processing step,avoiding its interference with subsequent detection,and using the SNIC superpixel algorithm to obtain more refined pre-segmentation results.Secondly,by introducing sea-sky-line estimation parameters,it is used to adjust the semantic segmentation model initialization and prior parameter information to improve the water surface segmentation accuracy of the entire model.Finally,the H-shaped boundary set is selected to construct a Gaussian background model in the water area,and the non-water pixels in the water surface are judged as obstacle targets.Experimental results show that this method can improve the accuracy of water surface obstacle detection under monocular conditions.(2)Aiming at the missed detection of obstacles targets caused by the blurring of the horizon near the sea-sky-line,an obstacle detection method based on the secondary segmentation model under binocular conditions is proposed.This method realizes obstacles detection through the secondary detection model and the three-dimensional verification link.Firstly,using a smoothing algorithm for preprocessing to reduce the interference of some noises in the image and enhance the edges and textures of obstacles.Secondly,by introducing the disparity map,the single-view model is extended to a joint stereo view semantic segmentation model and the improved segmentation model is combined to obtain the water area and improve the accuracy of water surface segmentation.Finally,the background-based saliency estimation method is used to realize obstacle detection,and the limit constraint and stereo matching method are used in the stereo verification link to verify the obtained obstacle detection results and eliminate the false detection caused by flashing and other reasons.Experimental results show that this method can improve the accuracy of water surface obstacle detection under binocular conditions.