Contents I. Introduction
Object tracking is an essential capability for a situational awareness system. It entails inferring information regarding objects of interest from noisy sensor data. Tracking multiple objects adds further complexity due to uncertainty in the association of measurements and number of objects [1]. Furthermore, if an object generates multiple measurements, extended object tracking methods can be used to estimate the shape and size, the extent, of the object in addition to its kinematic states. This provides more information, but necessitates more complex target models, further increasing the complexity of the problem [2]. To aid in tackling this complexity, one approach is to utilize information from targetprovided measurements. In the field of waterborne transport, such target-provided data already exist in the form of the Automatic Identification System (AIS), a crucial system for vessel navigation. It allows the vessels to exchange information with other vessels and nearby shore stations, to increase the situational awareness of the crew, and help them avoid collisions with other vessels. This system is widespread and mandated by law for ships above a certain size (300 gross tonnage) and all passenger ships [3]. Therefore, it has a key role to play in tracking such ships [4]. AIS has also been used for collision avoidance systems for trials of an autonomous surface vehicle (ASV) [5]. The data provided by AIS can be fused with information from other sensors to perform multi-object tracking and recent studies have shown that this enhances tracking performance across various multi-object tracking frameworks [6]–[9]. Recent work has demonstrated how to incorporate target-provided measurements in multiple extended object tracking [10] using an extended object Poisson multi-Bernoulli mixture (PMBM) filter with a Gaussian process (GP) target model [11]. In this work, the AIS message is also used to estimate the size and shape of an object in addition to the kinematic states of position and velocity, which improved the extent estimation and tracking performance. However, fusing AIS information in the context of extended object tracking gives rise to additional challenges. One such challenge is the fact that in reality, the reported position in the AIS message does not necessarily correspond to the center of the ship. At sea, this minor discrepancy is not an issue, since the distances between vessels are generally quite large. However, when a ship is close enough for extended object tracking to be relevant, this discrepancy should be taken into account, especially for larger vessels. A specific environment where this is relevant is inland waterways, which are relatively narrow compared to the size of the vessels, meaning that vessels will pass close to one another. In this paper, we build on the work in [10] to develop a multiple extended object tracking method using AIS and LiDAR data and use it to track large (CEMT Class II or above [12]) vessels in an inland waterway, with data collected near the Albert Canal in Belgium. The main contributions compared to [10] include the estimation of the offset between the reported AIS position and the estimated centroid of the extended object, a more specific initialization procedure for AIS measurements, and the demonstration of the method in a dataset that combines LiDAR data with real AIS data. The paper is organized as follows; in Section II we present the relevant background theory, in Section III we present the specific method and our contributions and in Section IV we present the results on the gathered data, compared to the benchmark methods.
