Twin network target tracking method based on bounding box positioning quality perception

• Main Authors: SHI YANG, WU LINGFAN, JIANG HUA, DING XUHUI, LU WEN, SHI WENJIE
• Format: Patent
• Language: Chinese; English
• Published: 30.01.2024
• Subjects: CALCULATING; COMPUTING; COUNTING; PHYSICS

The invention discloses a twin network target tracking method based on bounding box positioning quality perception, and the method comprises the steps: obtaining a to-be-tracked video, inputting the to-be-tracked video into a completely trained tracking model, and extracting a template feature and a search feature of a current video frame; the tracking model is obtained by training an initial tracking model by adopting a target loss function containing triple loss; according to the template features and the search features, regression feature vectors and classification feature vectors are obtained; the regression feature vector comprises the discrete probability of the parameter of each bounding box obtained by regression; the classification feature vector comprises the target confidence of the bounding box obtained by classification; determining a positioning quality score of the bounding box according to the regression feature vector, and performing weighted optimization on the target confidence according t