This paper proposes a fast and reliable method for anomaly detection and localization in video data showing crowded scenes. Time-efficient anomaly localization is an ongoing challenge and subject of this paper. We propose a cubic-patch-based method, characterised by a cascade of classifiers, which makes use of an advanced feature-learning approach. Our cascade of classifiers has two main stages. First, a light but deep 3D auto-encoder is used for early identification of “many” normal cubic patches. This deep network operates on small cubic patches as being the first stage, before carefully resizing the remaining candidates of interest, and evaluating those at the second stage using a more complex and deeper 3D convolutional neural network (CNN). We divide the deep auto-encoder and the CNN into multiple sub-stages, which operate as cascaded classifiers. Shallow layers of the cascaded deep networks (designed as Gaussian classifiers, acting as weak single-class classifiers) detect “simple” normal patches, such as background patches and more complex normal patches, are detected at deeper layers. It is shown that the proposed novel technique (a cascade of two cascaded classifiers) performs comparable to current top-performing detection and localization methods on standard benchmarks, but outperforms those in general with respect to required computation time.

A fast and accurate video anomaly detection and localisation method is presented. The speed and localisation accuracy are two ongoing challenges in real‐world anomaly detection. We introduce two novel cubic‐patch‐based anomaly detector where one works based on power of an auto‐encoder (AE) on reconstituting an input video patch and another one is based on the power of sparse representation of an input video patch. It is found that if an AE is efficiently trained on all normal patches, the anomaly patch in testing phase has a more reconstruction error than a normal patch. Also if a sparse AE is learned based on normal training patches, we expect that the given patch to AE is represented sparsely. If the representation is not enough sparse it is considered as a good candidate to be anomaly. For being more fast, these two detectors are combined as a cascade classifier. First, all small patches on test video frame are scanned, those which have not enough sparse representation are resized and sent to next detector for more careful evaluation. The experiment results show that the method mentioned here has a better performance especially in run‐time measure than state‐of‐the‐art methods on two UMN and UCSD benchmarks.

In this paper, we propose a method for real-time anomaly detection and localization in crowded scenes. Each video is defined as a set of non-overlapping cubic patches, and is described using two local and global descriptors. These descriptors capture the video properties from different aspects. By incorporating simple and cost-effective Gaussian classifiers, we can distinguish normal activities and anomalies in videos. The local and global features are based on structure similarity between adjacent patches and the features learned in an unsupervised way, using a sparse auto-encoder. Experimental results show that our algorithm is comparable to a state-of-the-art procedure on UCSD ped2 and UMN benchmarks, but even more time-efficient. The experiments confirm that our system can reliably detect and localize anomalies as soon as they happen in a video.

Explainability of deep neural networks has been receiving increasing attention with regard to auditability and trustworthiness purposes. Of the various post-hoc explainability approaches, rule extraction methods assist to understand the logic that underpins their functioning. Whereas the rule-based solutions are directly managed and understood by practitioners, the use of intervals or crisp values in the antecedents that rely on numerical values might not be intuitive enough. In this case, the benefits of a linguistic representation based on fuzzy sets/rules are straightforward, as these semantically meaningful components ease the model understanding. This paper proposes fuzzy rule-based explainer systems for deep neural networks. The algorithm learns a compact yet accurate set of fuzzy rules based on features' importance (i.e., attribution values) distilled from the trained networks. These systems can be used for both local and global explainability purposes. The evaluation results of different applications revealed that the fuzzy explainers maintained the fidelity and accuracy of the original deep neural networks while implying lower complexity and better comprehensibility.