Position estimation of multiple sea vessels using a stereo-based camera system and neural networks

Georgis, Stefanos; Γεώργης, Στέφανος

dc.contributor.author	Georgis, Stefanos	en
dc.contributor.author	Γεώργης, Στέφανος	el
dc.date.accessioned	2020-12-07T10:31:06Z
dc.date.available	2020-12-07T10:31:06Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/52334
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.20032
dc.description	Συστήματα αυτοματισμού	el
dc.rights	Default License
dc.subject	Sea-vessels-detection	en
dc.subject	Stereo	en
dc.subject	Neural networks	en
dc.subject	Deep learning	en
dc.subject	Position estimation	en
dc.title	Position estimation of multiple sea vessels using a stereo-based camera system and neural networks	en
heal.type	masterThesis
heal.classification	Naval engineer	en
heal.language	en
heal.access	free
heal.recordProvider	ntua	el
heal.publicationDate	2020-07-10
heal.abstract	In computer vision, triangulation via arranging two cameras in a stereo setup has become the norm in order to estimate the 3D pose of a particular object of interest and is used in most autonomous robots to help perceive the environment. In experiments limited to laboratory environments, classical computer vision techniques such as stereo correspondence search and triangulation work decently well. Moreover, one could utilize off-the-shelf equipment such as the Kinect sensor or Intel RealSense. The drawback being that such sensors have limitations when taken outdoor and have only a limited range (up to a few meters). This makes it infeasible to use such setups for long-range pose estimation. Range and time-of-flight sensors can be used to extract 3D information using raw data provided by such sensors from point clouds. But again, detecting particular objects in such point clouds is non-trivial. Having to do this for multiple objects of interest only compounds the task. Although, time-of-flight sensor manufacturers are trying to cut down costs and make such with competitive prices but are still a long way from manufacturing accurate sensors available at a competitive price such as cameras (which are orders of magnitude cheaper and provide most information per cent). In maritime industry the most famous surveillance systems and the most common are based on time-of-flights sensors like Radar. Although they are quite accurate most of the times they could not provide three-dimensional positional informations. To this end this thesis propose a low-latency real-time pipeline to detect and estimate 3D position of multiple Sea-Vessels using just two images of the same scene from a stereo based camera system. More specifically this thesis presents the design, development and implantation of a stereoscopic Sea-Vessels detection and localization system, aiming to provide accurate and robust results that could be used for avoidance collision in autonomous shipping. Our design goals are to provide a solution, which could be orders cheaper than the traditional solutions but providing more positional information for the detected objects, accurately and robustly. For this reason, after exhaustive search in current available hardware and perception algorithms technology we proposed a stereoscopic system exploiting neural networks for detecting Sea-Vessels. An improved fast horizon line detection pipeline is also presented and implemented in order to eliminate false Sea-Vessel detections. For estimating the 3D position of those detections, our system exploits the informations provided by the stereoscopic view. Furthermore, an improved 3D estimation algorithm is proposed, using just as measurements the detected Sea-Vessels in the current frame. This eliminate the need of precisely positioning specific markers in Ship’s hull and calibrate them with respect to the stereo rig. Our real time prototype system is capable of achieving 5FPS of continues detecting and pose estimating of Sea-Vessels in different Sea environments Finally the performance of the system is evaluated by conducting several tests in different lighting conditions, after the testing and approval of each sub-module of the system	en
heal.advisorName	Papalambrou, George	en
heal.committeeMemberName	Παπαδόπουλος, Χρήστος	el
heal.committeeMemberName	Γκίνης, Αλέξανδρος-Αλβέρτος	el
heal.academicPublisher	Σχολή Ναυπηγών Μηχανολόγων Μηχανικών	el
heal.academicPublisherID	ntua
heal.numberOfPages	101 σ.	el
heal.fullTextAvailability	false