Adaptive Target Classification with Imaging Sonar as a Partially Observable Markov Decision Process

摘要

Discriminating between different types of objects on the seabed using high-resolution imaging sonar is challenging, in part due to the inference of a 3-dimensional shape using one or more 2-dimensional projections. Often, more than one aspect of a detected object is required in order to correctly classify it as one of a number of targets of interest such as an influence mine, or as a benign non-target. When the sensor is equipped by an autonomous underwater vehicle (AUV), the acquired aspects are usually determined by a pre-planned mission rather than in response to the sensor data (i.e. the process is purely deliberative rather than reactive). In this paper, the multi-aspect target classification problem is modeled as a partially observable Markov decision process (POMDP). The solution to a POMDP is called a policy which provides a means for an AUV's control system to determine a course of action in response to the incoming sensor data, such as classifying an object as a target or not, or obtaining data from a specific aspect to reduce the uncertainty. The components of a POMDP for multi-aspect object classification with an AUV equipped with a sidescan are formulated. A numerical simulation is undertaken to assess the performance of the resulting policy and the vehicle path which is reacting to simulated sensor data is shown. The simulation was performed using two targets (a cylindrical object and a small, wedge shaped object) and a non-target class modeled as an elliptical object with uniformly distributed major and minor axes. The resulting policy was executed for many iterations and compared to one which obtains two perpendicular aspects. The classification accuracy of the POMDP model, as measured using a confusion matrix, was markedly superior to the cross-hatching strategy, while obtaining an average of 2.22 aspects (versus 2 for cross- hatching).