Analysis of the behavior of objects in potentially hazardous environments

GOUIZI, FATMA (2025) Analysis of the behavior of objects in potentially hazardous environments. Doctoral thesis, Faculté des Sciences et de la technologie.

Text gouizi_fatma_these_finale.pdf Download (23MB)

Abstract

Object behavior analysis systems represent one of the most effective approaches to enhancing security, ensuring smoothness, and promoting safety in public spaces, roadways, and hazardous environments such as intersections, level crossings, and pedestrian crossings. Among the various techniques employed in behavior analysis, anomaly detection stands out as a particularly significant method. However, it presents considerable challenges due to the complexity of video environments, difficulties associated with object detection, and the ambiguous definitions of various types of anomalies. This thesis explores behavior analysis in video, progressing from object detection to classifying these objects based on their states, ultimately determining whether they exhibit normal or disturbed behavior. Each process step is studied in detail, incorporating a comprehensive review of existing techniques, including both classical methods and those based on deep learning. The study also addresses various scientific and technical challenges and evaluation metrics. The proposed methodologies and the results obtained from experiments conducted on various datasets are presented at the end of the study. The primary goal of the initial step is to robustly detect and localize objects while considering environmental constraints, such as variations in lighting, weather conditions, and background movement. To achieve this, we introduce a novel strategy that utilizes a novel background subtraction architecture. Our network architecture is conceptualized as a nested network, which is based on residual autoencoder blocks featuring enhanced. This structure, referred to as "Nested-net," allows residual autoencoders to improve feature generalization by extracting multi-scale features at each level, thereby effectively addressing multiple challenges. The subsequent step focuses on detecting objects’ abnormal behaviors by introducing a new approach that employs a convolutional autoencoder to extract spatial and temporal representations from the appearance and motion of object patches. This is achieved through data transformation techniques aimed at enhancing feature learning and classification accuracy. Ultimately, the results obtained from various datasets demonstrate consistent performance, surpassing existing state-of-the-art techniques. This not only validates the proposed method’s effectiveness but also makes it a highly efficient and recommended solution for practical applications, thereby enhancing the safety and security of public spaces

Actions (login required)

View Item