Anomalous Action Recognition Research for Few-shot Learning

Abstract

The research of action recognition has achieved excellent progress and results, but the common datasets in action recognition research usually have a large number of samples and are suitable for the training of neural network, such as UCF-101 and Kinetics. For anomalous action research, it is difficult to collect and make datasets due to the rare occurrence of anomalous actions. Therefore, the research on few-shot learning of action recognition has certain significance. Based on the above, this paper studies the few-shot learning of anomalous action recognition. We present an anomalous action recognition network based on few-shot learning. We clip the videos in UCF-Crime datasets and other videos collected on the network to make a new dataset named UCF-Crime+. The action recognition network pre-trained by kinetics named I3D is used to extract the feature of the anomalous action video. We use a few-shot learning network, named prototypical network to classify the video feature samples, so as to realize the few-shot learning and classification of the anomalous video. In order to increase the difference between clusters, we use the triplet loss function. At the same time, the feature fusion of I3D network is carried out, which increases the utilization and fusion of deep and shallow features. The experimental results show that the improved few-shot learning model has higher accuracy than the base network.