Identifying and preventing abnormal behaviour in public areas has become more critical, as the reporting cases of such issues are growing nowadays. This emphasises the need to detect such abnormal situations in public spaces. Artificial Intelligence advancements based on Deep Learning and Computer Vision have greatly aided in the automated identification of anomalous occurrences. The proposed model offers a complete approach for detecting and analysing abnormal behaviour using an optimised YOLO network, enhanced through Adam Optimization, histogram equalisation, and other refinement techniques. In order to identify abnoraml behaviour precisely, the system employs optimisation methods, which also help to increase the detection process’s accuracy. The program also tracks observed items over time to find abnormal behaviour patterns. An improved YOLO framework is used for human detection, and the system’s performance is measured in terms of its capacity to recognise and analyse abnormal behaviour. The model proposed provides a remarkable accuracy of 99.46% in identifying and analysing abnormal behaviour, making it applicable to various real-time applications.
The rise in violence and crime has prompted the installation of surveillance cameras in public spaces to ensure people’s safety. However, the current system can only analyse past abnormal events with human assistance, making it impractical to identify such situations in real time. Therefore, there is a need for a real-time anomaly detection model. Such a model would significantly enhance the safety of the general population, particularly vulnerable groups such as the elderly, women, and children. Automatic Video Anomaly Detection can be employed to safeguard sensitive areas, including identifying enemy presence in military operations or protecting industrial resources and workers’ well-being. Similar to network security’s Intrusion Detection System, this video-based system can anticipate and identify intrusions and abnormal events ahead of time.
Detecting and analysing abnoraml behaviour is crucial in computer vision and security systems. Computer vision enables computers to interpret and analyse visual information from images or videos, making it essential for identifying unusual or suspicious behaviour in real-time video footage. Advanced deep learning and computer vision techniques have improved the accuracy and effectiveness of automated approaches for spotting abnormal behaviour. These systems employ motion analysis, shape analysis, and object identification techniques to identify and analyse abnormal behaviour patterns. Applications of abnormal behaviour detection systems include surveillance, traffic management, public safety, and healthcare monitoring. The progress of computer vision and deep learning has increased accuracy and efficiency in automated techniques for detecting abnormal behaviour.
However, optimising the detection performance of frameworks like YOLO (You Only Look Once) relies on factors such as the training dataset, optimisation method, and post-processing approaches. This research presents an optimised YOLO-based system for accurate and efficient detection and analysis of abnormal behaviour. The proposed system incorporates several modules, including optimized training of a Convolutional Neural Network (CNN) using Adam Optimization and pre-processing with Histogram Equalization (CDF method). For temporal analysis, it employs the Euclidean Distance Tracker for object tracking, while detection quality is improved through post-processing methods such as confidence thresholding and non-max suppression. Additionally, edge-based segmentation using the Prewitt operator and an enhanced YOLO framework are applied for accurate human detection.
Anomaly detection techniques in computer vision are vital for assisting law enforcement agencies in identifying and preventing criminal activities. Computer vision systems can detect suspicious behaviours like vandalism, theft, or violence by monitoring public spaces, transportation systems, and critical infrastructure. These systems enable early intervention, potentially preventing crimes or minimising the impact. The accuracy and effectiveness of automated systems for identifying abnormal behaviour have been greatly enhanced by recent advancements in computer vision and deep learning.
One notable framework is YOLO, which has shown promising results in various applications. However, optimising the performance of YOLO requires considerations such as the training dataset, optimisation method, and post-processing approaches. This research aims to present an optimised YOLO-based system for precise and real-time detection and analysis of abnormal behaviour.
Detecting and analysing abnormal behaviour in public places requires a multidisciplinary approach combining computer vision, data analytics, machine learning, and human behaviour psychology. This method leverages technological advancements and advanced algorithms to observe and analyse the behaviours and interactions of individuals in public places. Surveillance technologies, coupled with high-resolution sensors and advanced video analytics algorithms, enable detecting and analysing abnormal behaviour patterns. These algorithms leverage machine learning and computer vision techniques to identify suspicious activities, such as loitering in sensitive areas or unlawful entrance to restricted zones.
Analysing anomalous activity is crucial for maintaining safety and security across various domains. Some of the applications and significance of this research are:
Accurate and timely detection of abnormal behaviour is crucial for ensuring public safety and preventing potential risks.
Recent businesses focus more on customer-centric approaches, and gathering unbiased customer feedback has become a priority for marketing functions. As a result, studies using computer vision and facial feature emotion identification have become more popular, especially for marketing purposes. Computer vision is employed in health monitoring for structural infrastructures. The method utilises computer vision techniques to convert time series signals into grayscale images, mimicking human visual perception. The images created from the time series signals enable the deep neural networks to identify potential anomalies, such as structural damage or incorrect data, with high accuracy. A strategy based on big data for detecting abnoraml behaviour in healthcare settings was discussed. The approach uses association rules to analyse vast amounts of healthcare data and uncover patterns suggesting abnormal behaviour, contributing to better anomaly detection in healthcare systems.
The proposed model offers several contributions, including training and testing with CNN using Adams Optimization, histogram equalisation with the CDF method, Euclidean Distance Tracker, edge-based segmentation using the Prewitt Operator, and human detection using the Optimised YOLO Framework. This work aims to develop a reliable and efficient system for real-time detection and analysis of abnormal behaviour, which includes identifying criminal activity, safety hazards, and health emergencies. Real-time detection of abnormal behaviour is challenging due to the complexity and dynamics of the environment and the need to recognise subtle changes in behaviour patterns. The proposed system addresses these challenges by utilising an optimised YOLO network and various modules such as histogram equalisation, object tracking, non-max suppression, and edge-based segmentation to detect and analyse abnormal behaviour accurately. The system’s effectiveness is evaluated through experiments and analysis, demonstrating its high precision and performance.
Here is how the remainder of this project is supposed to go: The literature on abnormal behaviour detection methods is discussed in Sect. 2. The proposed fusion model is shown in Sect. 3. Section 4 deals with implementation, and Sect. 5 summarises the implementation results. Section 6 concludes with a discussion of the result and future work.
Afiq et al. explored various methods for diagnosing abnoraml behaviour in crowd scenarios, including feature-based approaches and deep learning strategies. Al-Dhamari et al. proposed a method of detecting unusual behaviour using transfer learning along with a binary support vector machine. In this approach, the model is initially trained using a source dataset and then further improved by training it on a target dataset to enhance its performance.
Arifoglu and Bouchachia introduced a Recurrent Neural Networks (RNNs) model to identify activities and detect abnormal behaviour by capturing temporal relationships. The training process involves a temporal understanding of behaviour patterns. In a subsequent study by Arifoglu and Bouchachia, they focused on using Convolutional Neural Networks (CNNs) to identify abnormal behaviour in dementia patients based on visual inputs.
Benmakrelouf et al. addressed the problem of identifying abnoraml behaviour in virtualised systems by translating resource-level measurements to service-level metrics. The model’s strategy aims to improve anomaly detection accuracy while minimising false positives. Cheng et al. introduced a state-of-the-art online behaviour anomaly detection method in Virtualized Network Function (VNF) services, employing machine learning techniques to enhance network security.
Cosar et al. focused on abnormal path and event detection in video surveillance, introducing a method based on trajectory modelling and clustering to identify abnormal behaviour in monitored scenes. Direkoglu proposed a method that uses motion information images and CNNs to detect abnormal crowd behaviour, extracting motion information from video frames and leveraging CNNs for classification.
Du et al. presented a method for detecting abnormal user behaviour using a selective clustering ensemble, combining multiple clustering algorithms to identify anomalies in user behaviour patterns. Fanta et al. introduced a Single-Tunnelled Gated Recurrent Unit (SiTGRU) for abnormality detection, which captures long-term dependencies and detects abnormal behaviour in sequential data. Fernando et al. developed a Long Short-Term Memory Networks (LSTM) framework through soft hardwired attention mechanisms for human route calculation and abnormal event detection, enabling the model to predict future trajectories and identify abnormal behaviours.
Recent studies have also highlighted the integration of deep learning, YOLO-based architectures, and advanced optimisation techniques in diverse domains. The study demonstrated the effectiveness of a YOLO Tiny architecture for multimodal fusion in lung disease classification, while explored blockchain-enabled monitoring systems for public works projects, showing the adaptability of AI-driven detection frameworks. Similarly, addressed anomaly detection in Ethereum smart contracts using integrated clustering, and applied hybrid CNNs for intelligent traffic prediction in smart cities. Choquette and Lei provided an overview of advancements presented in the SPIE proceedings, highlighting emerging trends in optical engineering and their potential relevance to surveillance and anomaly detection applications.
Energy efficiency is increasingly critical in real-time video analytics, especially for deployment on low-power devices such as surveillance nodes or IoT-enabled cameras. Recent research has emphasised the importance of lightweight architectures and computational profiling to achieve high detection accuracy with reduced resource consumption. Motivated by this, the proposed system aims not only to maximise detection accuracy but also to optimise efficiency through algorithmic refinements.
In addition to deep learning approaches, earlier research also explored statistical and model-based methods for abnormal behaviour detection. The study proposed a Markov Random Walk model to detect loitering behaviour in surveillance footage, showing that probabilistic models can effectively capture unusual pedestrian activity. Later, developed an integrated framework for detecting suspicious behaviours in video surveillance, highlighting the importance of combining motion features and contextual cues for robust detection. Related work also extended this line of research into broader applications, such as disaster event analysis in big data environments, demonstrating the adaptability of such frameworks to different domains. These contributions laid important groundwork for subsequent deep learning-based approaches, such as the proposed Optimised YOLO framework, which builds on these foundations by offering improved real-time performance and accuracy.
The problem statement is to design a system for detecting and analysing abnormal behaviour in real-time applications. The objective is to identify behaviour patterns that deviate from normal, such as unusual pedestrian movements or unexpected objects, in order to enhance safety and security in public spaces, workplaces, and private residences. For evaluation, the UCSD dataset is employed, which primarily focuses on anomalies like running, loitering, or the presence of non-pedestrian objects (e.g., bicycles, vehicles). While the dataset is limited to these categories, the proposed framework is designed to generalise to a broader set of abnormal behaviours, including violence, theft, and accidents, when applied to more diverse datasets. The system addresses the critical challenge of accurately detecting and localising abnormal behaviour in complex sequences with multiple objects and activities, demonstrating its potential in real-world scenarios.
