Acoustic image models for navigation with forward-looking sonars

Abstract

Cost and miniaturization of autonomous unmanned vehicles (AUV) drive component reuse and better sensor data analysis. One such component is the forward looking sonar (FLS) which can be used for obstacle avoidance and to extract vehicle state information. However, autonomous feature extraction of images from the FLS is difficult due to the noise inherent in the sensor and the sensor's susceptibility to interference from other acoustic devices. This thesis investigated techniques to detect and classify common acoustic noise artifacts and common objects in a single frame. Other techniques require three or more frames to filter objects from other noise sources. A combination of probabilistic and template-based models were used to successfully detect and classify acoustic noise and objects. One common noise source is the micro modem which was detected 100% of the time with 1% false positives. Objects such as the ocean floor were correctly classified more than 93% of the time in most sites. Due to the short development time frame, the software was developed with a two-stage approach. First, a high level scripting language was used for rapid prototyping of different classification techniques. In order to meet the time-constrained requirements of the target software, the classification algorithms were encapsulated as C++ classes in an object oriented design once the desired techniques were identified.