Aerial-ground person re-identification method via position-aware and local pre-interaction

In response to the issues of poor discriminative power in pedestrian features and insufficient model generalization in existing aerial-ground person re-identification methods, this paper proposes an aerial-ground person re-identification method via position-aware and local pre-interaction. A positionally biased self-attention mechanism is proposed, which incorporates positional information into attention scores to guide the model in focusing on tokens at key positions within the input sequence. This enhances the model's perception of spatial relationships between image patches and improves the discriminability and robustness of pedestrian features. A prompt-local pre-interaction module is created that uses a single, inexpensive pre-interaction to create fine-grained linkages between prompt semantics and local attributes. This strengthens the prompt vector's ability to perceive specific local details in the current view and enhances the model's capability for fine-grained detail reconstruction. The label smoothing regularization training strategy is introduced, converting original one-hot encoded hard labels into soft labels. This encourages the model to learn smoother and more generalizable feature representations, mitigates overfitting, and improves overall model performance and generalization. The validity of the proposed strategy is confirmed by extensive trials on the public aerial-ground person re-identification dataset LAGPeR, which show that it effectively enhances Re-Identification performance.