Verification of multiple model neural tracking filter with ship’s radar Witold Kazimierski, Grzegorz Zaniewicz, Andrzej Stateczny Faculty of Navigation, Chair of Geoinformatics Maritime University of Szczecin Szczecin, Poland
[email protected],
[email protected] Abstract— The paper presents the research focused on verification of multiple model neural filter in real time environment on board of the ship. The filter is a result of the works in the research project and is implemented as a software application. The application is then connected to real radar making the on-line and off-line verification possible. The filter was checked against IMO requirements and was compared to other commercially used filters. Estimation errors and delay was the basis for the verification. Keywords- radar tracking, marine radar, neural networks, GRNN
I.
INTRODUCTION
Determining of movement parameters of other vessels is the key issue for navigational officer on board of the marine ship. This information is the basis for any maneuver decision, so it should be as accurate and as actual as possible. One of the most important source of information about other vessel’s movement is radar target tracking widely spread at sea. For most of the vessels at sea it is mandatory to have a radar with target tracking function. The quality of target tracking depends on the tracking filter applied. Most commonly used are the numerical filters based on Kalman filtering, like Extended Kalman Filter (EKF) or multiple model filters. Another approach might be to use nonlinear methods like artificial neural networks. These have been studied by the authors and the results have been presented on International Radar Symposia since 2003. For last two years the concept of multiple model neural filter has been developed and this paper presents the results of one of the phases of verification research. The idea was to implement the filter for tracking targets with the real data gathered directly on board of the vessel. These have been made by software implementation of the filter and then by connecting PC to NMEA radar. Such a system allowed to perform research in two ways – off-line (with recorded data) and on-line. The on-line verification allowed to observe how the system performs in the real working environment. The off-line verification on the other hand allowed for analyzing filter settings and its influence on estimation in real environment. The paper presents the research itself and its results. First, IMO requirement for marine target tracking and the concept of
Research work financed by Polish Ministry of Science and Higher Education under the research project “Development of target tracking methods for marine radars with the use of neural filtering”
multiple model neural filter is presented, then the description of research set is proposed and finally the results of research experiment. TRACKING REQUIREMENTS FOR MARINE RADARS
II.
In 2004 IMO has issued a new set of requirements for target tracking, which are suitable for vessels buit after 1st of July of 2008. They are included in the Resolution of Maritime Safety Committee – MSC.192(79). The accuracy requirements are presented in table 1. They are defined for non-maneuvering vessel with 95% probability. TABLE I.
TRACKED TARGET ACCURACY (95% PROBABILITY FIGURES) ACCORDING TO IMO
Time of steady state
1
min:
movement
tendency
3
min:
movement
prediction
Relative course
11º
3º
Relative speed
1,5 kn or 10 %
0,8 kn or 1%
CPA
1,0 Nm
0,3 Nm
TCPA
−
0,5 min
True course
−
5º
True speed
−
0,5 kn or 1%
The requirements are defined for a non-maneuvering vessel with 95% probability. There are no requirements for maneuvering targets at all. The exact scenarios for testing of the equipment are included in IEC 62388. These are however not so important in this paper, as in real conditions the research do not use test scenarios. III.
MULTIPLE MODEL NEURAL FITER
The filter used in the research consists of several elementary neural filters based on General Regression Neural Network. The structure of such a filter was thoroughly presented in previous IRS’s. The results of previous research have shown that such a filters might be used for tracking of marine targets even with better results than Kalman filter. The best results have been however achieved with different filter parameters for different dynamic of target’s movement. Thus the idea of multiple model neural filter arose. Suitable mixing
methods have been proposed and filter parameters have been adjusted based on simulation research.
ARPA or may be shown on PC screen. Such a set allows to perform the research described in this paper.
GRNN itself is in fact neural implementation of kernel regression algorithms proposed in the nineties. GRNN computes weighted average of teaching vectors, which are observed movement vectors of the target. The weights are the values of Gaussian kernel function for the distances of input vector to teaching vector. Introducing of this filter allowed to use more than two observed past vectors, which gave promising results in the research. The switching module is to select one of the elementary filters and transmit it to the filter output.
The main goal of the research presented in this paper was to verify multiple model neural tracking filter on board of the vessel in real environment. Tracking errors and tracking delay were used as tracking quality indicators. The results were compared to IMO requirements and analyzed from the practical point of view. The research were performed in two stages – off-line and on-line.
The filter described above was implemented in ownprogrammed software with the use of MS Visual Studio and in such form was included in research set. IV.
RESEARCH SET
Maritime University of Szczecin is the owner of two research vessels. One of them – NAWIGATOR XXI is a 60 meters long ship designed for mostly for training purposes, but also for hydrographic surveys. She is fully equipped with navigational aids. One of them is marine radar JRC JMA-53209, which has been used for the research in the first option. In the second option another research vessel was used, called HYDROGRAF XXI. She is in fact a boat of 10 meters long used mostly for hydrographic survey in inland waters and for other research in the area of inland shipping. During the research project, HYDROGRAF XXI was equipped with broadband 3D radar by Navico with the indicator of Simrad, which has a so called mini ARPA implemented. The research set scheme in both cases is presented in fig. 1.
V.
THE RESEARCH
A. Research concept The off-line stage means that the real data from radar were first recorded while vessel was at sea and they were then played back for the research and for estimation with analyzed filter. This method was used with JRC radar. More than 40 tracks were recorded and a few of them were chosen for the research. This off-line verification allowed to survey the influence of filter settings for tracking quality. The on-line stage was being done with Broadband radar on board of HYDROGRAF XXI. The set was installed on board of the ship and the real time research has been provided. Both stages required collecting of movement information for the other ships from other sensors than radar. This included mostly AIS and VHF verification. B. Research scenarios Unlike in case of simulation in the real environment research one has no influence on the movement of target ships. Thus the idea of off-line research arose. From all the tracks recorded a few was chosen to provide a variety of situations. The main impact was however laid on course maneuvers. The scenarios can be basically divided into a few categories: •
vessel not maneuvering
•
slowly turning
•
dynamic but short course maneuver
•
large course alteration.
Precise description of researched situations will be included in full paper.
Figure 1. The research set
In such a proposal radar is supplied with power by ship’s network (in other solution it can be supplied with independent generator). Own ship’s data like position, speed and course are transmitted directly from ship’s sensors. For acquisition and target separation radar software is used. Target data are then send via NMEA to personal computer with the suitable software. The software receives data, encodes them from NMEA format and process them for tracking filter. The multiple model neural filter this then applied and the output of are estimated course and speed of target vessels. These are then coded back to NMEA format and may be entered as external
C. Research results The research results allowed to verify the quality of tracking with proposed multiple model neural filter. Final results will be presented in full paper. They will be showed in tables and in graphs. They will include comparison of different filters and to the IMO requirements as well. As an example fig. 2 is given, which presents an estimated course for a ferry STENA BALTICA during course alteration. The results for neural multiple model filter (in this case two-model) are presented in comparison to AIS, JRC radar and Kalman Filter. The results show that MMNF (Multiple Model Neural Filter) reacts faster for the maneuver and it follows the maneuver. The errors and the delay are smaller than for the other filters.
Figure 2. Experiment results – estimated course for course-altering ferry for different filters.
SUMMARY The paper will show the verification phase of the research focused on providing a new filtering method for marine radar target tracking. This research focused on the on-board use of the filter, however other research will also verify the method in other condition (like shore station). The results of the research suggests that a method presented might be worth of interest when comparing to other commercial solutions and in the aspect of IMO requirements. REFERENCES [1]
[2]
[3] [4]
[5]
[6]
[7] [8]
[9]
Y. Bar Shalom, X. R. Li, “Estimation with Applications to Tracking and Navigation: Theory Algorithms and Software”, John Wiley & Sons, Inc., NY USA, 2001 W. Juszkiewicz, A. Stateczny, “GRNN Cascade Neural Filter for Tracked Target Maneuver Estimation”, Neural Networks and Soft Computing, Zakopane 2000 T. Kantak, A. Stateczny, J. Urbański, “Basis of automation of navigation” (in polish). AMW, Gdynia 1988. W. Kazimierski, “Two – stage General Regression Neural network for radar target tracking”, Polish Journal of Environmental Studies, Vol. 17, No 3B, 2008. W. Kazimierski, “Selection of General Regression Neural Net-work’s Training Sequence in the process of Target Tracking in Maritime Navigational Radars”, Polish Journal of En-vironmental Studies, Vol 16A., 2007 X. R. Li, V. P. Jilkov, “A Survey of Maneuvering Target Tracking— Part V: Multiple-Model Methods”, IEEE Transactions on Aerospace and Electronic Eystems, Vol. 41, 2005. A. Stateczny (ed.), “Radar navigation”, GTN, Gdansk 2011 (in polish) A. Stateczny, W. Kazimierski, “General Regression Neural Network (GRNN) in the Process of Tracking a Maneuvering Target in ARPA Devices”, Proceedings of IRS 2005, Berlin 2005. W. Kazimierski, A. Stateczny “Adjusting multiple model neural filter for the needs of marine radar target tracking”, Proceedings IRS 2011, Leipzig 2011