 |
 |
 |
 |
| |
|||||
|
|
|
The following article describes a novel algorithm for the detection and classification of helicopters from radar signatures. Radar returns from helicopters contain distinct periodic components produced by the main rotor blades known as blade flashes as well as other components reflected from the body and tail rotor. These transient radar signatures are highly non-stationary and are functions of the helicopter speed and aspect relative to the radar. Classical approaches, based on decomposing the signal into discrete frequency components (Fourier analysis), do not work. MITRE has developed a robust solution to this problem using probabilistic signal models referred to as Hidden Markov Models (HMMs). A HMM is a collection of models that are used to approximate portions of the signal. Since the waveform is not stationary over time, a number of these models must be pieced together to approximate the entire waveform. The choice of model to use and the order in which the models are connected form the basis for the HMM approach. A problem of great interest to many military organizations is the ability to detect and classify targets from remotely sensed data. Helicopter detection and classification is an important automatic target-recognition problem that falls into this category. Most of us can tell when a helicopter is in the area from the distinctive pulsating sound the rotor blades produce. Radar signatures from helicopters also contain distinctive periodic components know as blade flashes. These transient radar signature components occur when the helicopter rotor blades are aligned with the radar antenna producing specular reflections. Helicopter target signatures obtained in this manner are difficult to classify using classical Fourier spectral estimation methods. Through conducted research in developing technology for the Dominant Battlespace Awareness project, MITRE has adapted pattern recognition methods used in state-of-the-art speech recognition systems to the problem of the helicopter radar signature detection and classification. The specific pattern recognition methods employed in this project use probabilistic signal models referred to HMMs to represent target signatures of interest. HMMs represent target signatures as sequences of elementary components. Both the structure of the elementary components and the order in which they can appear are statistically described in the HMM procedure. In an HMM, the elementary target signature components are represented by a set of statistical generation processes. Each statistical generation process is associated with a state of the HMM. An additional statistical selection mechanism determines which generation process (state) should be used to represent individual elements of the target signature. The specific structure of the selection mechanism for a given HMM is represented by a set of state transition probabilities. This nested statistical structure allows real-world target-signature variation and corrupting noise effects to be accurately represented, producing a robust and reliable method of signal detection and target classification. Differences in helicopter blade length, number of main rotor and tail rotor blades, blade width, and rotor frequencies produce distinctive radar signatures that can be used to discriminate between different types of helicopters, as well as between helicopters and other targets of interest. Using HMM methods, radar signatures can be analyzed on either a global basis, where the probability that a specific helicopter is present can be determined, or on a local basis, where individual time samples in a target signature can be related to events of interest, such as the presence of blade flashes. In the same way that the speech-recognition application quantified both nominal and speaker-dependent characteristics of words of interest, the HMM helicopter detection and classification algorithm used radar signatures, recorded over a wide range of operational conditions, and a training procedure to quantify helicopter radar signature characteristics. The standard HMM training process, the Baum-Welch algorithm, takes sets of signatures from known targets and determines optimal values for the observation probabilities and state transition probabilities via a maximum likelihood procedure. In the HMM helicopter detection and classification algorithm developed in this project, data about the rotor system geometry and operation characteristics of individual helicopters are used to "pre-train" the HMMs in the same way speech recognition systems use grammatical knowledge to assess the probability of various word sequences. After this "pre-training" was done, a Baum-Welch procedure was used to reveal features in the helicopter radar signatures related to the fine details of main rotor hub geometry and tail rotor characteristics. These additional features improved both the detection and classification performance of the HMM algorithm. It should also be noted that these additional details could not have been deduced from the training data via Baum-Welch procedure without using the information gained during the "pre-training" process. In addition the "pre-training" process constrained the HMM state transition probabilities in a manner that significantly improved the computational efficiency of both the Baum-Welch training and subsequent HMM analysis of radar target signatures. The HMM approach was evaluated using X-Band radar data from military helicopters recorded at Ft. A.P. Hill. After initial adaptive clutter suppression and blade-flash enhancement preprocessing, a set of approximately 1,000 raw in-phase and quadrature (I/Q) data records were analyzed using the HMM approach. A correct target classification rate that varied between 98 percent for a Pulse Repetition Frequency (PRF) of 10 KHz to 91 percent at a 2.5 KHz PRF was achieved. For more information, please contact Walter Kuklinski using the employee directory. |
Solutions That Make a Difference.® |
|
|
