Vehicle counting in drone images: An adaptive method with spatial attention and multiscale receptive fields

We propose an altitude-adaptive vehicle counting method with an attention mechanism and multiscale receptive fields that optimizes the measurement accuracy and inference latency of unmanned aerial vehicle (UAV) images. An attention mechanism is used to aggregate horizontal and vertical feature weights to enhance spatial information and suppress background noise. The UAV flight altitude and shooting depression angle are considered for scale division and image segmentation to avoid acquiring distance measurements. Based on the dilation rate, we introduce a receptive field selection strategy for the trained model to exhibit scale generalization without redundant calculations. A distribution-aware block loss is optimized via $k$ roots to balance the loss of sparse and crowded regions by dividing the density map. Experiments on three authoritative datasets demonstrate that compared with CSRNet, the proposed method improves the mean absolute error by 29.4%–54.0% and mean squared error by 28.6%–41.2% while reducing the inference latency. The proposed method exhibits higher counting accuracy than lightweight models including MCNN and MobileCount.