Wavelet Neural Network for Maneuvering Target Tracking

In maneuvering target tracking, the target motion model constitutes one of the fundamental elements. An effective maneuvering target model should accurately characterize various motion states during target maneuvers. Commonly utilized models include the Constant Velocity (CV) model, Constant Acceleration (CA) model, time-correlated Singer model, and the "Current" Statistical model for maneuvering targets. All these models employ maneuver frequency to represent target maneuver characteristics. In practical implementations, a fixed maneuver frequency is typically adopted. However, actual target maneuver durations continuously vary in real scenarios, meaning the maneuver frequency dynamically changes. Consequently, using a fixed maneuver frequency necessarily introduces estimation errors. When the sampling period falls between 0.5-2 seconds, tracking accuracy increases as maneuver frequency decreases. Nevertheless, the maneuver frequency remains fixed and cannot dynamically adapt to the target's true state. Therefore, this paper proposes a neural network-based adaptive adjustment method for maneuver frequency that enables dynamic frequency variation according to target maneuvers. This approach enhances state estimation accuracy and improves tracking performance. The algorithm employs wavelet neural networks for offline training, ensuring good real-time performance. By real-time modification of the maneuver frequency in the "Current" Statistical model, the algorithm can adaptively adjust the frequency parameter, bringing the tracking algorithm closer to the target's actual motion state. From an implementation perspective, the wavelet neural network would typically involve preprocessing layers for feature extraction from motion data, hidden layers with wavelet activation functions for temporal pattern recognition, and output layers generating optimized frequency parameters. The offline training phase would utilize historical maneuver data with backpropagation optimization, while real-time operation would involve feedforward computation with minimal computational overhead.