AI RESEARCH & DEVELOPMENT ANALYSIS

The literature on human-robot interactions identifies two main approaches: on one hand, studies focusing on extracting specific features to enable robots to approach and interact directly with humans, and on the other hand, studies centered on socially aware navigation, whose primary goal is to avoid interfering with human group dynamics. The first group of research focuses on how robots can extract and process individual features, such as facial expressions, gestures, and postures, to improve the quality of direct interaction. These works primarily use 2D images obtained from RGB cameras, as they strike a balance between precision and computational efficiency. For instance, Putro et al. [14] employ a deep neural network-based classifier to identify facial expressions in real time from 2D images, allowing the robot to adjust its behavior according to the user's emotions. Feature extraction using 2D methods remains the most common choice due to their ease of implementation and lower computational demands compared to 3D systems [8, 11, 17]. On the other hand, studies focused on socially aware navigation use more advanced methods to extract three-dimensional features, though these approaches are less common due to higher technical complexity and the use of costly sensors such as LiDAR or RGB-D cameras. In this domain, Silva et al. [15] propose a real-time social navigation framework for densely populated indoor environments, combining a depth camera (RGB-D)-based perception model and an indoor positioning system to detect social agents. Their approach includes a "Social Heatmap" that quantifies human density in the en-vironment and a multi-layer path planner to avoid congested areas, demonstrating effectiveness in complex scenarios such as hospi-tals. Girgin et al. [7] present a LiDAR-based system that generates 3D point clouds to estimate people's positions and trajectories in real time. Although these sensors are highly precise, their cost and computational demands limit broader adoption. Meanwhile, Kim et al. [10], while also focusing on robot navigation in social environ-ments, do not specify the use of 3D sensors, suggesting reliance on 2D methods or combinations of more accessible sensors. Despite advancements in three-dimensional sensors, 3D meth-ods are not yet standard in human-robot interaction applications, primarily due to their high cost, processing complexity, and greater computational demands. However, recent work such as Silva et al. [15] shows that it is possible to integrate affordable RGB-D sensors (e.g., the Intel® RealSense™ D435i) with adaptive plan-ning algorithms, achieving a balance between 3D precision and computational efficiency. In most current applications, 2D-based feature extraction remains dominant due to its ease of implementa-tion and effectiveness in many social interaction scenarios [7, 10]. Nevertheless, recent advances in detection technologies, such as YOLOv8, have narrowed the gap between 2D and 3D approaches. While YOLOv8 primarily operates on 2D images, its real-time de-tection capabilities and advanced segmentation make it particularly valuable for social robotics applications. When combined with 3D sensors like RGB-D cameras or LiDAR, this system enables robots to efficiently extract both planar and spatial features. These integrated detection and analysis capabilities are essential for robots that must simultaneously process individual interactions and navigate social spaces effectively. In this context, the use of RGB-D cameras, such as the Intel® RealSense™ D435i, represents an intermediate solution between 2D and 3D methods, offering precise 3D perception at a more accessible cost. This approach overcomes some limitations of more complex 3D sensors, such as LiDAR, while improving accuracy and adaptability in dynamic environments. Thus, the combination of accessible technologies and advanced algorithms, such as those proposed in this work, has the potential to bridge the gap between traditional approaches and the demands of more complex human-robot interaction applications.

In this section, we present our proposed solution to address the challenges of feature extraction in three-dimensional spaces. Our ap-proach utilizes the Intel® RealSense™™ Depth Camera D435i, which enables the precise generation of 3D point clouds and the sub-sequent extraction of keypoints. This technological integration facilitates advanced tasks such as group detection, people count-ing, centroid extraction for each individual, orientation estimation based on torso and head positions, and the identification of social interaction patterns and group-level descriptions. The implementation leverages the efficiency of YOLOv8 as a lightweight detection backbone, further fine-tuned through transfer learning specifically for shape classification in social groups. During inference, the system maintains minimal computational demands due to two key factors: (1) the Intel® RealSense™ D435i provides geometrically consistent 3D data through hardware-accelerated depth sensing, thereby eliminating the need for computationally intensive point cloud reconstruction, and (2) YOLOv8's optimized architecture ensures efficient and accurate keypoint extraction. By processing exclusively the detected keypoints rather than analyz-ing the entire point cloud, the system strikes an optimal balance between computational efficiency and accuracy. By integrating all these components, our solution empowers a robot to dynamically analyze its environment and make informed decisions to enhance its interaction with human groups. The fol-lowing subsections detail the architecture and implementation of the proposed methodology, highlighting the advantages of using a depth camera as a key tool for three-dimensional perception.

This section discusses the experimental results derived from a mul-tidimensional analysis of spatial and morphological parameters within the observed scene. Experiments were conducted in controlled indoor environments using groups of 3 to 4 individuals, with continuous measurements performed at distances ranging from 2 to 6 meters. Data was col-lected from multiple scenarios, recording at least 50 samples per configuration. Performance metrics such as group detection accu-racy, centroid depth error, and orientation estimation error (with corresponding standard deviations) were systematically computed. The analysis of the experimental results is structured into four key aspects: (i) group detection, which assesses the model's ability to accurately identify and cluster individuals into coherent groups; (ii) centroid depth estimation at individual and group levels, evaluat-ing depth positioning for both single individuals and entire groups; (iii) individual head and body orientation angle estimation, analyz-ing torso and head orientation relative to the depth plane under different spatial conditions; and (iv) group shape estimation, char-acterizing the overall geometric configuration of groups of varying sizes. By integrating these aspects, the proposed approach provides a detailed representation of spatial distribution, group morphology, and individual orientations in a three-dimensional environment. This analysis not only enables a deeper understanding of the con-figurations of social groups but also validates the robustness and accuracy of the model under various spatial conditions, reinforcing its applicability in real-world scenarios such as service robotics, healthcare, and rescue missions, where precise interaction with human groups is critical.

Our experimental results demonstrate the robustness and prac-tical viability of the proposed approach for robotic systems re-quiring precise 3D interaction capabilities with individuals and groups. By integrating Intel® RealSenseMD435i-generated spatial data with YOLOv8-based keypoint detection, the method achieves 21% higher group segmentation accuracy compared to traditional 2D approaches while maintaining 70% accuracy in torso orientation estimation across dynamic conditions. The framework successfully analyzes group configurations (L/C/I patterns) and social dynam-ics through cost-effective 3D perception, providing quantifiable insights into crowd behavior with metrics such as centroid disper-sion and torso angle estimation error (<40°). These advancements enhance human-robot engagement in service robotics, healthcare navigation, and rescue operations, where real-time interpretation of spatial interactions is critical. Future work will address current limitations in head orientation tracking through multi-modal sen-sor fusion and extend validation to outdoor environments with larger participant cohorts. This research lays a foundation for future developments, in-cluding the integration of GPS georeferencing for centimeter-level localization in unstructured environments, the analysis of body lan-guage through limb kinematics and attention pattern recognition, and the fusion of the proposed methodology with existing 2D group detection techniques to improve system robustness. While adaptive social robots show potential in healthcare and rescue operations, ethical concerns such as overreliance on automation and privacy risks must be proactively addressed. Our system mitigates these is-sues by avoiding biometric identifiers and focusing on anonymized spatial metrics - for instance, monitoring patient interactions in care facilities without storing personal data, aligning with GDPR and IEEE ethical AI guidelines. Future deployments will require interdisciplinary collaboration with ethicists and end-users to bal-ance technological efficacy with societal norms. These measures, combined with ongoing technical improvements, will expand the applicability of automated group behavior analysis across diverse domains.