Autonomous mobile robots have emerged as essential resources across several sectors of society, demonstrating significant impact and transformative potential. In industrial environments, these systems facilitate efficient material transport and engage in complex collaborative tasks with human operators and other robotic units. In transportation, advanced driver-assistance systems harness intelligent features to enhance safety and improve the driving experience. In domestic settings, service robots are increasingly deployed to assist with activities of daily living while also addressing psycho-social needs, particularly among elderly populations. Two fundamental aspects of these autonomous systems are their capabilities for navigation and human interaction within dynamic, complex, and anthropocentric environments. These capabilities require the integration of advanced perception and decision-making to ensure safety and effectiveness and leverage contextual awareness. Consequently, autonomous mobile robots must adapt to non-stationary conditions, account for human unpredictability, and adhere to both explicit rules and implicit social norms, all while operating under real-time constraints. Central to the navigation of autonomous mobile robots and human-robot interaction is the understanding and prediction of human behavior. This comprehension is fundamental for robots to make more informed decisions in shared spaces. In this research project, human behaviors are embedded in trajectory patterns, which can significantly affect the operation of mobile robots. By developing accurate and efficient models to classify and predict future trajectory patterns of other dynamic agents, autonomous systems can be integrated seamlessly into human-centric environments. The dynamics of these agents depend on the current activity or agent type. For example, in traffic scenarios, humans use high-level semantic concepts to categorize these heterogeneous agents, such as "cars," "pedestrians," or "cyclists. In such contexts, semantic categorization emerges as a powerful cue that significantly improves the accuracy of trajectory prediction. Hence, this research project focuses on understanding agent-based semantic cues that influence trajectory patterns in human-centered environments. These cues can be derived from human-given categories (e.g., in road scenarios, we have "pedestrians", "cars", "cyclists"), or more informative and meaningful categories can be learned through machine learning methods. The goal is to study this semantic categorization and its influence on tasks such as trajectory prediction and classification by using machine learning techniques.