The invention discloses an unmanned aerial vehicle dynamic target detection method in
extreme weather based on
deep learning. The method comprises the following steps: S1, collecting a video
stream and
atmospheric visibility to construct a continuous frame
image sequence; s2, inputting a physical
perception segmentation network based on SegFormer to generate a space-time attention
mask; s3, performing weighted suppression on the original image features by using the
mask, and outputting an enhanced feature image set; s4, inputting an improved SwinTransformerV2 model, and extracting target depth
time sequence semantic features through space-time dual coding and an environment
perception adaptive position coding mechanism; s5, calculating a spatial uncertainty
score by using a
Bayesian algorithm, calibrating confidence, and screening high-confidence features; and S6, performing multi-frame matching and track correction, and labeling a target bounding box and a motion track in real time. According to the invention, meteorological interference
noise is effectively filtered out, the target sensing capability under low
visibility is improved, and the robustness and accuracy of detection under a complex environment are ensured.