An Efficient Anomalous Action Recognition Model Based on Out-of-Distribution Detection

Abstract

Detecting anomalous data is very important for the security issues of machine learning. Misjudging anomalous data as normal data may cause serious consequences. For supervised machine learning methods, detecting anomalous data is a big challenge, because anomalous data may be very diverse and it is difficult to collect all possible anomalous data. In recent years, action recognition has been widely used in surveillance systems and home care systems. The recognition of anomalous actions has also become an important requirement of the action recognition system. In this paper, we apply the method that has successfully detected anomalous data on 2D images to identify anomalous actions in videos. The proposed approach can directly identify anomalous actions as long as we train on normal action data. The experimental results show that the proposed approach has achieved significant improvements on anomalous action recognition.