Bird target tracking method and system based on target detection and motion prediction fusion

Abstract

The invention discloses a bird target tracking method and system based on target detection and motion prediction fusion. The method comprises the following steps: acquiring an initial frame image and an initial bounding box, and generating a feature template through a trunk tracking network; obtaining a current frame image, inputting the current frame image into a trunk tracking network, determining a search area based on a tracking result of a previous frame, matching the search area with the feature template, and outputting a preliminary prediction bounding box and confidence; inputting the current frame image into a motion predictor to generate a motion prediction bounding box; judging whether a triggering condition is met or not and generating an optimization bounding box; calculating the intersection-to-union ratio of the motion prediction bounding box and the optimization bounding box, and obtaining a final tracking bounding box through scene stability analysis; and updating the final tracking bounding box to the backbone tracking network and the motion predictor. By adopting the method, the tracking accuracy and robustness of the birds in a complex motion scene can be improved.

Images