Robust Moving Object Detection In Complex Videos Using Deep Learning Techniques

Abstract

newlinevi newlineABSTRACT newlineMoving object detection is a fundamental yet challenging task in the field of computer vision. It involves the identification and tracking of objects in motion within a video sequence. Despite significant advancements in the field, achieving human-like visual perception remains a complex problem for automated systems. The difficulty arises primarily due to the variability in object motion, scene complexity, occlusion, and environmental conditions. Furthermore, no single detection approach proves universally effective across all types of motion, as the speed of moving objects can vary significantly within video content. newlineTo address this, videos are typically categorized into three motion-based types: fast, moderate, and slow-moving object videos. Each category presents unique challenges. Fast-moving objects often result in motion blur and are difficult to track accurately due to limited frame rates in standard video capture. For instance, tracking a baseball in real-time during a game exemplifies the complexity of fast-motion scenarios. Moderate-moving objects may experience intermittent occlusion by background elements or other moving objects. A practical example includes monitoring vehicles at intersections for traffic analysis and safety compliance. In contrast, slow-moving object such as a person speaking in a video conferencing or compression test scenario pose a different set of challenges. Although movement is minimal, precise segmentation, especially of the full human body across frames, becomes more difficult due to subtle motion and appearance similarities with the background. newlineThis thesis investigates the detection of moving objects referred to as Video Object Frames (VOFs) across all three motion categories. It introduces an advanced deep learning framework designed to address the diverse challenges encountered in detecting slow, moderate, and fast-moving objects. The proposed system incorporates modified encoder-decoder architecture, leveraging transfer learning strategies and an enhanced Feature Pooling Framework (FPF) for multi-scale spatial and temporal feature extraction. This enables the system to capture both fine-grained details of slow motion and broader patterns of rapid object transitions. To further increase model robustness and prevent overfitting, a dynamic spatial dropout mechanism is integrated, allowing better generalization to unseen scenarios and diverse object speeds. newlinevii newlineAdditionally, the framework is evaluated using thermal video datasets, where detection relies on heat signatures instead of visible light. This application is especially critical for environments with poor visibility, such as nighttime surveillance, foggy roads, or smoky industrial areas. Thermal-based detection enables reliable identification of moving objects where traditional RGB-based methods often fail. newlineWhile various methods have been proposed over the years, most still fall short when applied to real-world scenes with dynamic object velocities and complex environments. This thesis aims to bridge that gap by proposing a unified, scalable, and robust architecture capable of effective object detection across different motion types and challenging video conditions. newlineIn this thesis, a deep learning-based framework is proposed for the detection of moving objects across varying motion dynamics in complex video scenes. Initially, the VGG-16 architecture is employed as the encoder backbone, integrated with a modified FPF module, to effectively capture and identify slow-moving objects. To enhance the model s capability across a broader range of motion, the improved FPF is further combined with the VGG-16 encoder to accurately detect slow, moderate, and fast-moving objects under diverse and challenging scene conditions. In pursuit of improved accuracy and deeper feature representation, the VGG-19 model is also utilized as an encoder, followed by the integration of ResNet-101 and the modified FPF. This combination leverages the complementary strengths of deeper networks and refined multi-scale feature fusion, resulting in a robust detection framework capable of handling complex motion patterns and cluttered backgrounds. newlineKeywords: Video surveillance, Background subtraction, Deep learning architecture, Convolutional Neural Networks, Pre-trained model, Transfer learning, Feature pooling framework, Contrast normalization, Leaky ReLu, Fine-tuning, Feature maps, Multi-scale detection, Precision, Recall, Precision, Accuracy, F1 Score, Intersection over union, Robustness, Real-time application.