The invention provides an unmanned aerial vehicle-mounted vehicle converse running detection method based on lightweight key point detection. The method comprises the following steps: step 1, constructing a
data set; five-point labeling of a central point, vehicle width and height and a rotation angle of a UAV-ROD
data set needs to be converted into four-point labeling of a bounding box central point, bounding box width and height and a normalized key point; 2, designing a model; the
network architecture of the YOLO11 is optimized, and a
light detection head LADH is adopted to replace an original detection head to form a YOLO11-LADH model; 3, vehicle
direction detection and retrograde driving judgment are carried out; and training the YOLO-LADH model on the UAV-ROD key
point data set, inputting the pictures of the
view angle of the unmanned aerial vehicle into the model to obtain four key points of the vehicle, and judging the driving direction of the vehicle through the obtained four key points. Through the lightweight detection head LADH, the detection precision and the reasoning speed are improved while the
model parameter quantity and the calculation quantity are reduced, the anti-jamming capability is high, judgment is accurate, multi-frame association is not needed, and an efficient solution is provided for real-time
traffic management.