The invention discloses a double-person interactive gazing
behavior recognition and mode conversion positioning method, which comprises the following steps of: 1, specifying the head positions of to-be-detected double target persons in an input video, and respectively acquiring a gazing point
heat distribution map, a gazing
direction vector and a drop point
label sequence through a gazing point
heat map generator; and step 2, using a deep feature extractor of a cross attention mechanism, and fusing related information to extract
gaze behavior features. 3, designing a multi-scale
gaze mode recognizer, extracting
gaze behavior
feature relevance on multiple time scales, and obtaining a gaze
behavior recognition result after decoding; 4, designing a mode conversion locator, learning inter-
frame difference characteristics, and obtaining the probability of a conversion point at each moment; and a fifth step of constructing a
loss function in combination with double tasks of double-person watching
behavior recognition and mode conversion positioning, and completing training of the constructed
deep learning network. According to the method, through multi-scale
time sequence feature fusion and a double-task joint training strategy, excellent performance is shown in double-person watching behavior recognition and mode conversion positioning tasks.