Tracking

The objective of preprocessing is to threshold raw radar data to obtain detections that will be used for track update. The aim of thresholding is to detect almost all of the target originated data points, while minimizing the number of false alarms. AUG Signals has made significant contributions in the area of automatic CFAR detection; the company is a world leader in CFAR research. Our CFAR detection technologies employing K-PDF or GC-PDF (Model-based CFAR), fusing multiple CFAR detectors (Multi-CFAR) and Markov chain CFAR detector outperform the closest competition by increasing probability of detection by 20% while maintaining a constant false alarm rate. In addition, AUG Signals has developed a novel detection algorithm that uses feedback from tracker in the form of predicted track positions and attributes, which consistently outperforms the state of the art.

AUG Signals has developed superior tracking algorithm that improves the tracking performance metric (e.g., track accuracy, track completeness, track consistency) by up to 40% compared to the state of the art.

Multi-layered approach

For robust performance in the presence of high clutter, (or high track density) tracks are divided into two categories; confirmed tracks and initial tracks. Confirmed tracks are validated while initial tracks are yet to be validated. AUG Signals applies a robust decision mechanism for track initialization, confirmation and deletion, which is based on track length, number of misdetections and the degree of association between the track and measurements. In addition, a three-layer association approach is used to associate a set of measurements with existing tracks and generate new tracks. This approach gives higher precedence to the targets in which the tracker has higher confidence.

Measurement-to-track Association

AUG Signals uses multidimensional data association between a set of tracks and multiple sets of measurements to significantly improve data association performance and, consequently, tracking performance. The multidimensional association is sequentially performed by m-best 2-D association algorithm. This algorithm borrows from Multiple Hypothesis Tracking (MHT) concept which refines various hypotheses as more data is available. Our association algorithm significantly reduces the computational load of MHT while maintaining similar level of performance, and is used for both active and passive measurements. Our track initiation unit generates new tracks using measurements that are not associated with currently tracked targets. While measurements from single passive sensor are used to generate angle only tracks, those from multiple passive sensors are used to obtain tracks in the Cartesian coordinate system. The state of the art of multiple passive sensor based tracking uses multidimensional association algorithm, which requires the full association tree. AUG Signals’ passive sensor based tracker associates measurements directly with tracks, which drastically reduces the association tree size and hence, the computational complexity (up to a factor of 10). In addition, there is up to 20% improvement in data association performance.

Tracking filter design

AUG Signals has significant experience in designing state of the art tracking filters for various applications. Interacting Multiple Model (IMM) filters based on Kalman models, with modes tuned to both linear motions and maneuvers, have been designed for tracking highly maneuvering targets. AUG Signals has developed an efficient algorithm to choose the IMM modes given the target types present in any scenario. For scenarios with nonlinear state transition and/or nonlinear measurement (e.g., ESM sensors) the Kalman filter is replaced by extended Kalman filter. In systems with high degree of nonlinearity Unscented Kalman Filter (UKF) is used. In addition, AUG Signals has significant experience in particle filter based tracker design.

In recent years one of the main emphases of tracking system development has been target recognition. The traditional approach for recognition, known as joint tracking and classification, updates the classification probabilities based on feature information (e.g., measurement amplitude, transmission wavelength, maneuver characteristics, etc.). Feature information is also used in track-to-measurement association. The main shortcoming of this approach is it does not generate unique ID for each target; rather, it assigns the targets to broader classes. Hence, target identity is not preserved when association ambiguity causes track-switch. AUG Signals has developed a novel target identification algorithm which creates and updates ID for each track that evolves separately from track state. Hence, in case of track switch, ID can be preserved, giving a complete picture of evolution of kinematic and feature information that leads to robust target recognition.

Multiple sensors with overlapping surveillance region provide redundancy (similar sensors) and complimentary view (dissimilar sensors). Efficient registration and fusion of sensor information is required to achieve this. AUG Signals has significant experience in multi-sensor registration and fusion applications (e.g., hyperspectral signal fusion, video-based superesolution reconstruction, and multi-sensor fusion for liquid analysis) spanning several fields. AUG Signals has expertise in two different approaches of tracking related multi-sensor fusion: centralized and distributed fusion. Centralized fusion is employed when information from all sensors is available in a central processing center. Alternatively, in distributed fusion, processed information (e.g., tracks) from sensors is fused, which saves communication bandwidth between distant sensors. AUG Signals applies robust sequential processing approach and track de-correlation approach in centralized and distributed fusion, respectively, to achieve optimal tracking performance.