This Small Business Innovation Research Phase I project will demonstrate the feasibility and effectiveness of novel nonlinear filtering and joint decision and estimation methods for robust and persistent tracking of multiple targets with nonlinearities in measurement and dynamic model. The key innovations in this effort are: i) a computationally efficient and accurate nonlinear point estimator that outperforms traditional extended Kalman filter and unscented filter ; ii) an approximate density estimation filter based on a numerical solution of Fokker-Planck equation with grid adaptation that provides an efficient alternative to particle filter based approaches, iii) a novel joint density and (point) estimation (JDE) framework that is optimal in information theoretic sense and enables efficient JDE for multi-target tracking, for example, joint detection and tracking. The project will benefit from University of New Orleans expertise in nonlinear filtering, target tracking, and statistical inference, and ObjectVideos ongoing research activities on target detection, multi-target tracking, efficient multi-frame data association, and distributed MTI tracking. The Phase I effort will include: development of proposed nonlinear filtering and joint decision and estimation technologies, development of performance models for nonlinear multi-target tracking, quantitative and qualitative evaluation of the proposed technologies, and demonstration of proof of concept.
Benefit: Automatic interpretation of sensory data has been a persistent topic of research in the areas of information theory, sensor fusion, computer vision, pattern recognition, machine learning, and psychology, but is still in a very primitive state. Robust detection and persistent tracking of targets in the scene is a critical first step towards enabling systems capable to interpret activities in the scene and provide timely situational awareness and effective forensic analysis capabilities to the analysts. The technologies proposed here enable persistent tracking of multiple targets with nonlinear dynamic and observation models. The proposed tracking techniques are generic and can be applied to many ISR sensors and applications that are characterized by complex models and uncertainties in the data, for example, urban surveillance from WAMI sensors, GMTI and AMTI tracking, 2D/3D particle tracking, and dynamic analysis of cellular organism in microbiology domain.
The proposed technologies can also be used to enable real-time and forensic analysis tasks that find a lot of applications in the domains of geospatial intelligence (GEOINT), persistent surveillance, and video analysis. These benefits include:
- Timely Situational Awareness: Timely availability of reliable intelligence is critical to support military missions and objectives. Knowledge regarding the trajectories of targets, their interactions with each other and other scene elements is a critical part of intelligence and its automated extraction from surveillance data will help expedite the dissemination of relevant intelligence to the commanders in the field.
- Reliable Inference: Existing automated video surveillance and monitoring tools allow users to specify security rules in the form of tripwires and exclusion zones on maps or images. Persistent tracking of targets will enable users to receive alerts with high precision thus increasing the reliability of the overall system and user-confidence on the output of the intelligent system.
- Comprehensive Video Analysis: Robust target tracking in videos is a step closer towards comprehensive video analysis, i.e., automatic extraction of a description of the elements, activities, and events in the scene.
Keywords: Nonlinear Filtering, Density Estimation, Target Tracking, Joint Decision And Estimation
