Learning Spatio-Temporal Features for Detecting Anomalies in Videos using Convolutional Autoencoder

Abstract

Automatic video surveillance systems are a recurrent topic in recent video analysis research. Anomaly detection is an interesting way for tackling this problem, because video analysis is tedious and exhaustive for humans. Depending on the application field, anomalies can present different characteristics and challenges for pattern representation, requiring the design of hand-crafted features (such as spatial and temporal information). Deep learning methods have achieved the state-of-the-art performance for many recognition problems in recent years and may be an interesting choice for learning features automatically, since it captures the 2D structure in image sequences during the learning process. The deep Convolutional Autoencoder (CAE) may be an interesting approach for anomaly detection since they can learn signatures automatically in an unsupervised way. This work purposes the use of a deep CAE in the anomaly detection context for learning spatio-temporal signatures from raw video frames. Similar our previous work, we use as anomaly score a successful strategy based on the reconstruction error of a package of frames. The proposed methods were evaluated by means of several experiments with public-domain datasets. The promising results support further research in this area.

Explainable Anomaly Detection in Videos Based on the Description of Atomic Actions

“…Previous approaches for the anomaly detection task have used hand-crafted features such as motion and appearance [6]. In the last years, Deep Learning (DL) techniques have achieved promising results in anomaly detection by learning better features with superior discriminatory power for video and images representation [7], [8]. DL techniques have also been applied to solve different tasks in computer vision, including action recognition [9] person re-identification [10], age and gender recognition [11], and clothing segmentation [12], [13].…”