Human Abnormal Behavior Detection Based on Multimodal Data Fusion
Xiangyang She, Zhiqi Xu · 2024
When the number of people increases or occlusion occurs, the human abnormal behavior detection method based on multimodal data fusion has low reliability of human behavior data information, low detection accuracy, and poor adaptability. This article used a multimodal data fusion model to solve the above problems. Mature human pose recognition technology can be used to extract human joint point data from videos. Joint coordinates can be transformed into angle and distance features of human behavior to express human posture. Machine learning methods can be applied to analyze and process joint features, obtaining data distribution features that are conducive to identifying abnormal movements. By comparing the multimodal data fusion model with other mature algorithm models, the study found that under the two evaluation methods of X-sub and X-view, the recognition accuracy of the multimodal data fusion model was 88.7% and 97.8%, respectively, which was 10.1% and 42.8% higher than the baseline method. This article identified 8 types of behaviors, among which 5 types of behaviors had an accuracy rate of over 80%, and the accuracy rate of behavior detection had all reached over 65%. Compared to other model methods, it has better practicality.