Inert Mine Object The goal of this research is to use advances in machine learning and computer vision on the ever improving acoustic sensors for object detection and classification in underwater applications. A specific goal of the research will be object detection and classification of mines found on the sea floor. Specifically we will focus on two mine types from a 900 khz Marine Sonic side-scan sonar from a REMUS 100 Autonomous Underwater Vehicle.

The new and interesting element of this research will be experimentation with multiple computer vision features in a multi feature cascasde training method. The idea is to provide a pool of Side-scan sonar image of mine features that show promise for the application and let the AdaBoost feature selection process choose the most valuable features. One interested new feature, which is unique to sonar images, we call the shadow feature. Effectively it captures all information about shadows in the current window, including width, height, etc. and more interestingly, the altitude of the vehicle and the lateral location of the shadow in the image.

The other interesting focus of the research will be in adaptive learning techniques intended to alter the classification EEG Cap algorithms as new data becomes available in order to improve results in changing environments. One technique based on online boosting will be a batch update technique that considers all relevant information from previous training and combines with all of the new training data. Another interesting adaptive learning consideration will involve smart EEG caps that allow for a human in the loop to review images at high speeds and automatically determine which images were of interest to the operator.

Update on October 01, 2012

REMUS 100 deployed by marines The target AUV for this research is the Hydroid REMUS 100. This vehicle is currently used by the Navy for missions to find mines on the ocean floor. All of the data captured from these missions is analyzed post mission by trained Marines. Some of the data sets for the research come from this platform.

Chris collecting data from AUV

I will directly interact with the REMUS 100 for data collection and testing of output algorithms. As part of this research we have the lofty goal of moving outside of the general academic setting into an application or engineering setting by actually implementing our algorithm on a REMUS 100 and testing the capability in San Diego Bay.

The REMUS 100 uses a pair of 900 khz Marine Sonics side-scan sonars to capture and create the imagery of the sea floor. The resulting Marine Sonic Tiff (MST) files are generally post processed by proprietary software on a toughbook in the field. We will be augmenting this flow with our automatic target recognition and human in the loop approaches.

Update on October 01, 2012

Implement a solution for automatically detecting mine-like objects in side-scan sonar images, leveraging the recent increase in sonar image resolution and the advances in computer vision features and machine learning for object recognition. This merges research from: oceanic engineering, vision, and machine learning.

Update on October 01, 2012

Haar-Like Features:

Haar-Like Cascade results   The ROC for haar-like feature

Speeded Up Robust Features:

SURF testing example   The ROC for SURF feature

Shadow Features:

Shadow processing example   The ROC for Shadow feature

Update on October 01, 2012