A Target Identification Method Based on Uncertainty Estimation and Evidence Theory
Leping Lin, Jinwei Chen, Xiaodong Cai, Li Jin · 2022
Aiming at the problem that the accuracy of target identity recognition is easily affected by environment, target state and hardware limitations in surveillance scenes, a target identity recognition method based on uncertainty estimation and evidence theory is proposed, which fuses face and pedestrian information. Firstly, an uncertainty estimation module based on Dirichlet distribution is proposed to realize the reflection of the influence degree of various complex factors on single-mode characteristics in a monitoring scene by numerically simulating and expressing the subjective logic of single-mode decision. Secondly, a combination decision-making method based on DS (Dempster-Shafer) evidence theory is proposed to realize multi-modal dynamic fusion decision-making and eliminate the possible decision-making conflicts among single modes. Finally, a Bayesian risk-based classification loss method is proposed to realize the generation of evidence-level information under multi-modality through constraint network, and the test is carried out on the data set under the urban surveillance scene, and the accuracy rate reaches 87.87%, which is superior to the single-modality recognition and splicing fusion method.