Target Re-Identification Architecture for Low-Altitude Micro UAVs Based on Siamese Neural Networks and Metric Learning

Small unmanned aerial vehicles (sUAVs) are widely used in security surveillance, target tracking, and intelligent transportation. However, traditional target recognition methods are susceptible to environmental complexities such as illumination variations, viewpoint differences, and occlusions, leading to degraded performance. Target re-identification technology enables the matching of the same target across different spatial and temporal conditions, making it a key approach for enhancing UAV-based target recognition capabilities. This paper investigates a target re-identification method based on a Siamese neural network and metric learning, constructing a comprehensive system framework and proposing an adaptive recognition process for dynamic environments. An optimized Siamese neural network architecture is designed, integrating improved feature extraction and loss function strategies to enhance target discrimination and recognition accuracy. The model's convergence and robustness are validated through experiments, while metric learning techniques are employed for target matching. By incorporating cosine distance and TriHard Loss during training, the cross-device recognition performance is significantly improved. Experimental results demonstrate that the proposed method effectively enhances the accuracy and stability of target re-identification in complex environments, providing technical support for UAV vision perception systems.