UAV Landing Tracking With Lightweight Twin Networks and Kalman Search

The tracking algorithm is segmented into two independent networks: a template network for extracting target features and a search network for detecting targets in search regions. This segmentation allows each network to be optimized independently, reducing overall computational complexity while maintaining tracking precision through specialized feature extraction and detection mechanisms.

The patent replaces traditional filtering-based tracking mechanisms with a deep learning-based twin network system. The template network and search network utilize convolutional neural networks to extract and match deep image features, substituting mechanical filtering operations with intelligent feature-based matching that achieves both real-time performance and high precision.

The template network extracts target features offline from a predefined target template before the tracking process begins. This preliminary feature extraction avoids the need for continuous online template updates during tracking, preventing error accumulation while maintaining real-time performance through efficient feature matching in the search network.

The search region is dynamically adjusted based on target detection results and movement predictions. When the target is successfully detected, the search region is concentrated around the detected position. When the target disappears transiently, the search region is expanded to increase the probability of re-detection, thereby improving tracking reliability without excessive complexity.

Different feature channels are assigned different weights based on their effectiveness for specific target types. The system learns channel-specific weights through training, allowing important channels to contribute more to the matching process while less important channels contribute less. This differentiated feature utilization improves tracking precision for different target types without requiring complex manual feature selection.