Improving Multiple-Model Context-Aided Tracking through an Autocorrelation Approach

Abstract

This paper continues a previous work, where the context-aided tracker "ConTracker" was used to detect suspicious behaviors in maritime vehicle trajectories. ConTracker takes into account map-based contextual information - which includes water depth, shipping channels and areas/buildings with a high strategy value - to determine anomalies in ship trajectories. The different areas act as repellers or attractors that modify the expected trajectory of the tracked vessel. In the original scheme, a multiple-model adaptive estimator (MMAE) is used to estimate the noise parameters of the tracking system: sudden increases on the output reflect unexpected maneuvers - such as entering a forbidden area - that are translated as alarms. The work presented here shows the results obtained by implementing a generalized version of the multiple-model adaptive estimator (GMMAE). While the former approach uses information of the last cycle to update the weight/importance of each model, our proposal calculates a likelihood value based on the time-domain autocorrelation function of the last few indicators. GMMAE provides a much faster response, which ultimately leads to a general performance boost: alarms are faster and clearer. Compared with previous works, GMMAE is particularly effective returning back to normal state after an alarm has been raised: this results in alarms with a better defined duration. Results are presented over several simulated trajectories, featuring a variety of realistic anomalies which are correctly identified. They include direct comparison with the previous approach, for an objective demonstration of the achieved improvement.