Weakly supervised video anomaly detection is generally formulated as a multiple instance learning (MIL) problem, where an anomaly detector learns to generate frame-level anomaly scores under the supervision of MIL-based video-level classification. However, most previous works suffer from two drawbacks: 1) they lack ability to model temporal relationships between video segments and 2) they cannot extract sufficient discriminative features to separate normal and anomalous snippets. In this article, we develop a weakly supervised temporal discriminative (WSTD) paradigm, that aims to leverage both temporal relation and feature discrimination to mitigate the above drawbacks. To this end, we propose a transformer-styled temporal feature aggregator (TTFA) and a self-guided discriminative feature encoder (SDFE). Specifically, TTFA captures multiple types of temporal relationships between video snippets from different feature subspaces, while SDFE enhances the discriminative powers of features by clustering normal snippets and maximizing the separability between anomalous snippets and normal centers in embedding space. Experimental results on three public benchmarks indicate that WSTD outperforms state-of-the-art unsupervised and weakly supervised methods, which verifies the superiority of the proposed method.

Falls are a major health threat for older people. A timely assistance can reduce the extent of physical injury caused by the falls. Currently, low-cost and convenient video surveillance systems based on ordinary RGB cameras are widely used for improving the safety of people. The fall detection is a research hotspot in intelligent video surveillance. In this work, we propose an unsupervised fall detection method. The proposed method first converts the RGB video frames into human pose images to eliminate the background interferences and focus on human motion and protect privacy. Afterwards, the future pose images are predicted by using the continuous historical human pose images based on a constrained generative adversarial network (GAN). Finally, the prediction errors of the human pose images and the anomaly scores of actual poses calculated by using the traditional hand-crafted features are used to realize the fall detection. As compared to the existing vision-based fall detection methods, the proposed method possesses strong generalization ability, and is robust to environmental interferences and small local occlusions, and effectively protects the privacy, and avoids time-consuming data annotations. In addition, in this work, a new large-scale and comprehensive fall dataset is created and is available for download. We perform extensive experiments on the public benchmark datasets and the proposed dataset. The results demonstrate the validity and superiority of the proposed method.