Surface Ship Location Based on Active Sonar Image Data

A human observer can locate surface ships in sequences of active sonar images based on intensity features due to the ship’s hull, wake, and its emitted cavitation noise. Three different sector-scan sonar image target detection and tracking algorithms are examined here that will inform further research towards the goal of developing a computer algorithm to perform the surface ship detection. Introduction Active sonar emits an acoustic pulse into a body of water and collects signals from an array of hydrophones (underwater microphones) in a finite time after the pulse. Given a constant speed of sound, the range to a given target can be calculated from the time it takes the emitted acoustic pulse to travel from the sonar to the target and back, divided by 2, and multiplied by the speed of sound in water. Match filtering the hydrophone data with the the transmitted pulse increases the time resolution, and thus the range resolution, that can be extracted from the signal data containing reflections of the transmitted pulse. Beamforming “points” the response of the hydrophone array in a specific direction, or usually, many specific directions. Such conventional signal processing produces an bearing versus range intensity response image per emitted pulse. From emitted pulse to the beamformed result, the cycle for one emitted pulse is referred to as a “ping,” for historical reasons (reference?). In addition to the acoustic reflections, the hydrophones also receive other acoustic energy, including cavitation noise produced by a ship’s propeller. Despite a signal processing chain tuned to detect reflections of a transmitted pulse, this energy shows up in the resulting images. A human observer can visually locate surface ships in the bearing vs. range sonar images images by a combination of their characteristics. An acoustic reflection from the the hull of a Fig. 1. Portion of active sonar image showing radial cavitation noise spoke and reflection from boat wake. Source: ARL:UT stationary, silent ship becomes an intensity point in the sonar image. However, if the ship is moving under it’s own power, such a point is swamped by other features. An acoustic reflection from the persistent air bubbles in the ship's wake appears as an intensity line according to the wake’s position. Constant acoustic noise produced by the ship appears as a radial line in the image according to the ship’s bearing. See Figure 1 for an example of an active sonar image with a wake reflection and a noise spoke (beam vs. time plotted as azimuth vs. distance). A variety of applications could benefit from a computer algorithm that could automatically locate surface ships from sequential active sonar data. In particular, underwater vehicles must be able to avoid collisions with ships when surfacing. This requirement is well illustrated by the disastrous February 2001 collision between the USS Greenville nuclear submarine and a Japanese fishing boat. An autonomous underwater vehicle does not have the luxury of a human operator, but must still avoid such collisions whether surfacing or docking with a larger ship or submarine. Much work has been published on active sonar image feature tracking. Much of it is devoted to features that appear on the bottom or in the water column, and not to the specific problem of avoiding surface ships. However, many of the concepts studied may be applicable to the problem at hand. In general, target tracking from sonar data has two main parts. First is a filtering stage to refine the sensor data and extract target candidates, and second is a correlation stage to associate target candidates with a track. A matched filter and beamformer, as described above, will assumed to be the first part of the filtering stage for the purposes of this paper. The implementation of the rest of the system varies widely depending on methodology and application. Optical Flow Method: Chantler, et. al. devised a method that works over multiple pings to to separate stationary targets from moving ones, and then calculate the optical flow of the moving targets [1-2]. First, a 1-D Fast Fourier Transform (FFT) is applied to the the time sequence of each image pixel across multiple pings. A band pass filter and inverse FFT is applied to obtain images containing the dynamic targets, and a low pass filter and inverse FFT is applied to obtain an image with the static targets (See Figure 2). Each image is thresholded to obtain binary images containing distinct objects. The apparent motion of the brightness patterns, or optical flow, of the objects in the dynamic images is used to segment the image into significant objects and provide motion information about those objects. These observations are used to find possible associated objects in the next image frame; a tracking tree is constructed by progression through multiple frames and recording multiple possible paths. For each branch of the tree (a possible track), a compatibility measure is computed based on the expected position of the object in the next frame. A cumulative compatibility measure, or confidence value, is recomputed for each branch of the tracking tree with each new ping, where the maximum value is reported as the object’s actual track. Experiments were done with scuba divers moving among a group of pier legs, demonstrating low confidence measures assigned to tracked noise, and high confidence assigned to the divers for a fixed position sonar, a fast sonar ping repetition rate, and smooth target motion. LANE et al.: ROBUST TRACKING OF MULTIPLE OBJECTS IN SECTOR-SCAN SONAR IMAGE SEQUENCES 33 Fig. 3. Processing stages for moving object detection, motion characterization, and tracking. Fig. 4. FFT processing to separate moving and static observations in a sonar image sequence. Model-based tracking algorithms [21]–[23] are well suited for polyhedral and manufactured objects where a wire frame model exists. In these methods, 3-D polyhedral models of the objects are given. Detection and segmentation of the moving target hus reduces to a problem of recognition, which for sonar observations requires the motion information as a classification feature [3]–[6].