Self-Supervised Feature Extraction for Video Surveillance Anomaly Detection

• Davi D. de Paula UNESP
• Denis H. P. Salvadeo UNESP
• Lucas B. Silva UNESP
• Uemerson P. Junior UNESP

Resumo

The recent studies on Video Surveillance Anomaly Detection focus only on the training methodology, utilizing pre-extracted feature vectors from videos. They give little attention to methodologies for feature extraction, which could enhance the final anomaly detection quality. Thus, this work presents a self-supervised methodology named Self-Supervised Object-Centric (SSOC) for extracting features from the relationship between objects in videos. To achieve this, a pretext task is employed to predict the future position and appearance of a reference object based on a set of past frames. The Deep Learning-based model used in the pretext task is then fine-tuned on Weak Supervised datasets for the downstream task, using the Multiple Instance Learning training strategy, with the goal of detecting anomalies in the videos. In the best case scenario, the results demonstrate an increase of 3.1% in AUC on the UCF Crime dataset and an increase of 2.8% in AUC on the CamNuvem dataset.