面向异常行为检测的知识蒸馏算法研究与实现

In recent years, with the development of monitoring technology and the increasing emphasis on public safety, the demand for detecting abnormal behaviors of pedestrians in public places has been growing. Abnormal behavior detection models are usually deployed on edge devices, facing the challenge of limited storage resources. In response, existing methods mostly adopt model lightweighting techniques to reduce the resource consumption of models. However, these methods usually rely on a large amount of data for training, and the scale of abnormal behavior datasets is small with an imbalanced distribution of positive and negative samples, resulting in a decrease in the accuracy of lightweight models. In addition, there are correlations among abnormal behavior detection tasks. Different tasks share similar feature extraction processes, generating a large amount of redundant calculations. To solve the above problems, this paper studies knowledge distillation algorithms for abnormal behavior detection and proposes a cross - task adaptive distillation algorithm. This algorithm introduces a dynamic weight allocation strategy and an adaptive temperature adjustment mechanism, which improves the detection accuracy and generalization ability of lightweight models. The experimental results show that compared with common existing lightweight methods, the algorithm proposed in this paper achieves better lightweighting effects on abnormal behavior datasets.