SBIR-STTR Award

Performance of robotic systems would be enhanced through artificial intelligent fusion of the automatic target recognition (ATR) results (and associated underlying features ) of either a single platform or multiple platforms viewing the target/object from different position, i.e., multi-look fusion. Such fusion would be particularly relevant for partially obscured and/or background blended targets. We propose a multi-layer perceptron neural net for implementing the multi-look fusion. This neural net would be similar to that recently used by other ARL ATR researchers. Performance would be further enhanced by smart fusion of individual sensors classifier outputs, either on the same or multiple platforms, i.e., multi-sensor fusion. Our approach would start with the fusion scheme being investigated by us under another SBIR for Target Acquisition/Target Recognition (TATR) encompassing man-in-the-loop decision making. For the ARL SBIR, we will enhance the TATR fusion scheme by building an artificial intelligent agent that augments and automates aided fusion function, e.g., adaptive thresholding for multi-sensor fusion. This intelligent agent would also incorporate production rules for associating different views of the same object for multi-look fusion. the resulting algorithms will be applicable to ATR fusion implementations both on vehicles and fusion stations. Our multi-look and multi-sensor classifier fusion approaches are very practical for operational scenarios, because neither depends on precise co-registration of the various disparate sensors.

Keywords:
cooperatvie target acquisition multi-look classifier multi-sensor neural net

I-MATH proposes algorithm and software development for Robotic Perception using an Artificial Intelligence Enhanced Information Processing (AIEIP) system consisting of several innovative components: evolutionary (genetic) algorithms, neural nets, and expert rules with automated learning. The AIEIP system melds together an Intelligent Assistant (IA) learning system with I-MATH geometric hashing and piecewise level fusion classifiers. The IA eases demands upon the operator by using a dynamic window-based interface. The IA combines disparate sensors and algorithms into reliable and accurate operator information. This system is built upon ten years of I-MATH image/signal processing development experience. In particular, I-MATH geometric hashing algorithms were the only recognition software deemed to be sufficiently mature for incorporation into the operational Demo II vehicles, for the recently completed DARPA Unmanned Ground Vehicle program.

Benefits:
AIEIP will extend this capability by providing situation awareness using multiple sensor types and multiple UGV platforms. Regions and targets of interest can be much more thoroughly understood through the simultaneous and integrated multilook processing capability. The image processing pattern recognition algorithms sort through extensive scene information to locate and identify potential targets of interest. I-MATH has successfully demonstrated relevant technology with midwave and longwave FLIR, intensity and range LADAR, and MMW high resolution range radar sensors. Phase II is a watershed opportunity where for the first time all the disparate technique developments can be integrated into a single optimized algorithm architecture. The AIEIP II SBIR provides this opportunity because all the needed data bases are for the first time already in hand, validated, and truthed. Under this SBIR, I-MATH will deliver a complete set of software which would satisfy the TACOM UGV Demo III RSTA capability needed in FY00. Also, I-MATH has been contacted by the German Ministry of Defense and Dornier Aerospace for application of our algorithms/software to their new, telerobotic Weasel vehicle, particularly for processing its LADAR sensor imagery. Other applications include UAV's facility surveillance and the Intelligent Vehicle Highway System.

Keywords:
artificial intelligence evolutionary algorithms robotics expert systems neural nets