Event recognition and anomaly detection using machine learning

• Main Author: Peng, Xinggan
• Other Authors: Lin Zhiping
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science
• Online Access: https://hdl.handle.net/10356/173509
• Institution: Nanyang Technological University

Abnormal event recognition and anomaly detection aim to discover incidents, patterns or objects that do not conform to expected behaviour. In this thesis, our focus includes two parts: “illegal parking” activity detection tasks from images and anomaly detection with applications to time series data. For the first work of this thesis, a novel voting-based real-time illegal parking detection algorithm using images from in-vehicle cameras is proposed to achieve benchmark results for illegal parking detection tasks. The proposed algorithm can produce detection results with detailed illegal parking offences’ types. It is suitable for real-world dynamic scenarios without complex and high-cost installation procedures for data collection. To the best of our knowledge, our proposed algorithm is the first research work to achieve such functionalities. A novel image labelling method named “minimal illegal units” is introduced to link the vehicle and essential parking information and reduce labelling labour and time cost. Experiment results show that the proposed algorithm can detect illegal parking activities with multiple types of illegal parking offences and is robust to changes in working conditions. For the second and third works of this thesis, anomaly detection with applications to time series data is achieved based on machine learning and deep learning algorithms, respectively. Firstly, a machine learning framework named ELM-MI with DKS is proposed to detect anomalies based on mutual information estimation. The proposed dynamic kernel selection method by hierarchical clustering on unsupervised training data overcomes the limitations of the origin ELM-MI and better exploit temporal context to detect anomalies of various types. Extensive comparison experiments on the public and our collected datasets validate that the proposed framework is an effective solution for anomaly detection without large computational resource requirements. Finally, a deep learning framework named TCF-Trans is proposed to perform anomaly detection with applications to time series data using temporal context fusion. The proposed feature fusion decoder in the framework fuses features extracted from shallow and deep decoder layers to prevent the decoder from missing unusual anomaly details while maintaining robustness from noises inside the data. Meanwhile, the proposed temporal context fusion module exploits temporal context information of the data by a learnable weight to adaptively fuse the generated auxiliary predictions. Extensive experiments on public and collected transportation datasets validate that the proposed framework is effective for anomaly detection in time series compared with other recently proposed methods. In addition, a series of parameter sensitivity experiments and the ablation study show that the proposed method maintains high performance under various experimental settings.