SBIR-STTR Award

We propose a model-based temporal integration approach for automatic target recognition (ATR) using a single video stream. The video ATR problem is addressed as a verification and identification task. A generalized Hausdorff metric will be used for target to model matching and target to target tracking and verification. The Hausdorff metric is more robust to noise and allows a confidence interpretation for target recognition/verification. Our system can address stationary and moving target ATR using videos acquired from stationary as well as moving platforms. When fully developed, the system will integrate camera motion compensation, target detection, target tracking, target pose estimation, target verification, and target fingerprinting components. In Phase I, we will develop robust algorithms for 3D target pose estimation and sequential target recognition using a generalized Hausdorff metric. We will test and refine our existing algorithms for sensor motion estimation, target detection and tracking, and best view selection for video ATR applications. We will develop techniques for reliable and descriptive vehicle fingerprint extraction. The algorithms will be evaluated using videos acquired using a turn table as well as surveillance videos. During Phase II, we will develop, demonstrate and evaluate a prototype system capable of robust video ATR. The primary military applications for the proposed work are in target recognition from unmanned moving platforms. Commercial applications are in face recognition from a single video, and tracking a cargo laden truck or a ship over a long period of time

ImageCorp, Inc. proposes to design and develop a deployable video automatic target recognition (ATR) system for vehicle classification and recognition using a single video stream. The system will be built by improving the performance of the existing Phase I prototype system which was based on fusing frame-based decisions made either by a support vector machine or a Hausdorff metric based classifier and did not achieve full temporal integration of video frames. Experiments with real video streams illustrated many remaining challenges. Three major areas have been identified for further development and system integration in Phase II: (1) refining the Phase I software by adding capabilities for shadow detection, pose estimation, detection and tracking of multiple vehicles under occlusions, and simultaneous tracking and verification using particle filters; (2) investigating the utilization of CAD models to generate 3D fingerprints, enable classification of a much larger class of vehicles containing subtle variations, and assist target mensuration; and (3) addressing system integration issues related to real-time classification. Optimal video camera configuration will be studied, and system performance will be comprehensively evaluated. Two demonstrations will be performed using the deployed prototype system operating in realistic outdoor scenarios using live videos.

Keywords:
Tracking And Verification, Model Guided Recognition, Svm Classification, Hausdorff Metric, Decision Tree Classifier, Particle Filtering, Fingerprintin