The invention provides a construction operation pole falling
video detection method and device based on TPF-Net, and belongs to the field of
computer vision, and the method is characterized in that the method comprises the steps: carrying out the preprocessing of an obtained construction operation
site monitoring video or an unmanned aerial vehicle
aerial video, and carrying out the video frame extraction, size normalization, pixel
standardization and
noise suppression; and the stability of input data is ensured. One
branch is a
time sequence visual
branch, modeling is carried out on a video
frame sequence by adopting the
time sequence visual
feature extraction network,
global time sequence features are extracted, and RTFM loss is introduced, so that the sensitivity of a model to abnormal events such as inclination evolution and sudden
instability is enhanced; and the other
branch is an attitude
estimation branch, extracting pole
tower key point coordinates by using an attitude key point detection network, calculating a pole body inclination angle and forming a
time sequence angle sequence. Dimension raising is carried out on the angle features through a multi-layer
perceptron, weighted fusion is carried out on the angle features and the visual features in the same dimension space, unified feature representation is obtained, and finally a classification head is input to complete recognition of a
normal state and a pole reversing state. The method is suitable for rapid deployment and control and edge end real-time operation in the construction operation period, detection precision and calculation efficiency are both considered, and missing detection and
false detection of electric pole inclination and falling events under complex background, shielding and auxiliary supporting structure interference can be effectively reduced; and important reference is provided for safety management and control, graded early warning and emergency disposal of a construction operation site.