Feature-aided Random Set Tracking on a Road Constrained

This paper describes the application of finite set statistics (FISST) to a real-time multiple target road constrained feature-aided tracking problem. A vehicle of interest traverses the road network while other confuser vehicles cross paths with this vehicle. Features extracted from sensors are used to disambiguate the vehicle of interest from the confuser vehicles. The FISST formalism naturally leads to understanding ambiguity in the identity of targets. 1.0 Introduction Many tracking applications address the issue of monitoring one (or a few) vehicles of interest (VOI) in a background of confuser vehicles. One such program is the DARPA Affordable Moving Surface Target Engagement (AMSTE) program [1, 2]. Monitoring vehicles traversing a dense road network poses several complications since target state vectors can change rapidly due to target maneuvers at intersections [1]. There is often an association ambiguity between the measurement and the target positions because multiple roads and targets are in the sensor’s field of view. Adding false alarms and missed detection further increase the complexity of road tracking. Since these confuser vehicles can cross paths with the VOI, ambiguity in the kinematic state of the VOI is unavoidable [2, 3]. Disambiguating the VOI from the confuser vehicle requires an ID process that verifies the position of the VOI using features that are unique to the VOI. Traditional Kalman filter based approaches using multiple target tracking methods have difficultly in kinematic tracking of a VOI in a background of confuser vehicles because of limitations in the association process. The association process makes one or a few hard associations between targets and detections [4]. If incorrect, these hard associations lead to a greater certainty in target position than the data actually supports. The result is a false continuity in the track that implies known target IDs. This overly optimistic tracking result must be corrected by an algorithm introduced on top of the filter that accounts for confusion between tracks. We have introduced and tested a random set tracker (RST) that naturally avoids making these associations in purely kinematic applications [5]. The random set approaches do not explicitly define tracks and avoid the association ambiguity by statistically weighing all possible hypotheses and associations [5, 6,7]. As a result, it does not have the implicit track continuity that leads to difficulties in defining track ambiguity. By Kronecker producting the kinematic space with a feature space, the RST incorporates ID into tracking problems without modifying the framework or introducing patches on top of an inadequate framework. Similar to its tracking capabilities, the RST framework naturally accounts for ID ambiguity by avoiding hard associations between feature vectors and specific tracks. There are two basic feature-aided tracking applications that the current fusion algorithm can accommodate both applications [4]: 1) The simplest is the tracking vehicles of various classes, pick-up trucks versus sedans. In this application, the features of the various classes are predefined. Given a sensor measurement, the probability of a vehicle belonging to a specific class can be determined. 2) Another application requires that the system acquire features that uniquely identify the vehicle, such as high range resolution radar profiles [2]. After these features are acquired, the tracking application reduces to the classification problem with the assumption that only one vehicle belongs to each class. Determining when and how to acquire features is an inherent resource management issue. This paper outlines utilization of incoming sensor measurements to maintain tracks without explicitly addressing the issue of sensor management. Once the fusion framework is developed, it will be possible to predict future system behaviors, which will allow future development of resource management algorithms. The proposed RST implementation represents information in a framework that accommodates resource management algorithms performing both feature-aided tracking applications. The specific application we discuss in this paper assumes the existence of two classes of targets, a VOI class and confuser class. As in the AMSTE program, we will assume the VOI class contains a unique target that is defined by its feature vector (FV), but unlike AMSTE, the FV of the VOI is known a priori. The FV space we consider is a continuous vector space, such as length or color,