Recognizing human interactions in RGB videos is a critical task in computer vision, with applications in video surveillance. Existing deep learning-based architectures have achieved strong results, but are computationally intensive, sensitive to video resolution changes and often fail in crowded scenes. We propose a novel hybrid system that is computationally efficient, robust to degraded video quality and able to filter out irrelevant individuals, making it suitable for real-life use. The system leverages multi-modal handcrafted features for interaction representation and a deep learning classifier for capturing complex dependencies. Using Mask R-CNN and YOLO11-Pose, we extract grayscale silhouettes and keypoint coordinates of interacting individuals, while filtering out irrelevant individuals using a proposed algorithm. From these, we extract silhouette-based features (local ternary pattern and histogram of optical flow) and keypoint-based features (distances, angles and velocities) that capture distinct spatial and temporal information. A Bidirectional Long Short-Term Memory network (BiLSTM) then classifies the interactions. Extensive experiments on the UT Interaction, SBU Kinect Interaction and the ISR-UOL 3D social activity datasets demonstrate that our system achieves competitive accuracy. They also validate the effectiveness of the chosen features and classifier, along with the proposed system’s computational efficiency and robustness to occlusion.

Human object detection and recognition is essential for elderly monitoring and assisted living however, models relying solely on pose or scene context often struggle in cluttered or visually ambiguous settings. To address this, we present SCENET-3D, a transformer-driven multimodal framework that unifies human-centric skeleton features with scene-object semantics for intelligent robotic vision through a three-stage pipeline. In the first stage, scene analysis, rich geometric and texture descriptors are extracted from RGB frames, including surface-normal histograms, angles between neighboring normals, Zernike moments, directional standard deviation, and Gabor-filter responses. In the second stage, scene-object analysis, non-human objects are segmented and represented using local feature descriptors and complementary surface-normal information. In the third stage, human-pose estimation, silhouettes are processed through an enhanced MoveNet to obtain 2D anatomical keypoints, which are fused with depth information and converted into RGB-based point clouds to construct pseudo-3D skeletons. Features from all three stages are fused and fed in a transformer encoder with multi-head attention to resolve visually similar activities. Experiments on UCLA (95.8%), ETRI-Activity3D (89.4%), and CAD-120 (91.2%) demonstrate that combining pseudo-3D skeletons with rich scene-object fusion significantly improves generalizable activity recognition, enabling safer elderly care, natural human–robot interaction, and robust context-aware robotic perception in real-world environments.

Human activity recognition (HAR) is crucial in fields like robotics, surveillance, and healthcare, enabling systems to understand and respond to human actions. Current models often struggle with complex datasets, making accurate recognition challenging. This study proposes a quantum-integrated Convolutional Neural Network (QI-CNN) to enhance HAR performance. The traditional models demonstrate weak performance in transferring learned knowledge between diverse complex data collections, including D3D-HOI and Sysu 3D HOI. HAR requires better extraction models and techniques that must address current challenges to achieve improved accuracy and scalability. The model aims to enhance HAR task performance by combining quantum computing components with classical CNN approaches. The framework begins with bilateral filter (BF) enhancement of images and then implements multi-object tracking (MOT) in conjunction with felzenszwalb superpixel segmentation for object detection and segmentation. The watershed algorithm refines the united superpixels to create more accurate object boundary definitions. The model combination of histogram of oriented gradients (HoG) and Global Image Semantic Texture (GIST) descriptors alongside a new approach to extract 23-joint keypoints by employing relative joint angles and joint proximity measures. A fuzzy optimization process optimizes features that originated from the extraction phase. Our approach achieves 93.02% accuracy on the D3D-HOI dataset and 97.38% on the Sysu 3D HOI dataset Our approach achieves 93.02% accuracy on the D3D-HOI dataset and 97.38% on the Sysu 3D HOI dataset. Averaging across all classes, the proposed model yields 93.3% precision, 92.6% recall, 92.3% F1-score, 89.1% specificity, an False Positive Rate (FPR) of 10.9% and a mean log-loss of 0.134 on the D3D-HOI dataset, while on the Sysu 3D HOI dataset the corresponding values are 98.4% precision, 98.6% recall, 98.4% F1-score, 99.0% specificity, 1.0% FPR and a log-loss of 0.058. These results demonstrate that the quantum integrated CNN significantly improves feature extraction and model optimisation.

Group activity recognition in public environments is challenging due to dynamic formations, complex inter-person interactions, and frequent occlusions. Existing methods often emphasize individual actions, overlooking collective behavioral patterns. This work introduces a multi-modal framework integrating silhouette-based appearance and skeleton-based pose information for robust recognition in surveillance scenarios. You Only Look Once v11 (YOLOv11) detects persons, Segmenting Objects by LOcations version 2 (SOLOv2) segments instances, and AlphaPose extracts skeletons, followed by hierarchical grouping to form spatially coherent clusters. A hybrid feature extraction strategy combines handcrafted descriptors (Extended GIST (ExGIST), Distance Transform, Binary Robust Independent Elementary Features (BRIEF), Ridge) with deep representations, fused via multi-head attention. Feature selection is refined through a three-stage pipeline of Kernel Principal Component Analysis (K-PCA), mutual information ranking, and genetic algorithm-based optimization. Spatio-Temporal Graph Convolution Networks (ST-GCN) models spatio-temporal dependencies, while Long Short-Term Memory (LSTM) captures long-term dynamics for activity classification. On the Collective Activity Dataset (CAD), the framework achieves 96.80% accuracy, surpassing state-of-the-art approaches. Its modular design ensures scalability and adaptability for intelligent surveillance and smart city applications.

Inertial Sensor-based Daily Activity Recognition (IS-DAR) requires adaptable, data-efficient methods for effective multi-sensor use. This study presents an advanced detection system using body-worn sensors to accurately recognize activities. A structured pipeline enhances IS-DAR by applying signal preprocessing, feature extraction and optimization, followed by classification. Before segmentation, a Chebyshev filter removes noise, and Blackman windowing improves signal representation. Discriminative features—Gaussian Mixture Model (GMM) with Mel-Frequency Cepstral Coefficients (MFCC), spectral entropy, quaternion-based features, and Gammatone Cepstral Coefficients (GCC)—are fused to expand the feature space. Unlike existing approaches, the proposed IS-DAR system uniquely integrates diverse handcrafted features using a novel fusion strategy combined with Bayesian-based optimization, enabling a more accurate and generalized activity recognition. The key contribution lies in the joint optimization and fusion of features via Bayesian-based subset selection, resulting in a compact and highly discriminative feature representation. These features are then fed into a Convolutional Neural Network (CNN) to effectively detect spatial-temporal patterns in activity signals. Testing on two public datasets—IM-WSHA and ENABL3S—achieved accuracy levels of 93.0% and 92.0%, respectively. The integration of advanced feature extraction methods with fusion and optimization techniques significantly enhanced detection performance, surpassing traditional methods. The obtained results establish the effectiveness of the proposed IS-DAR system for deployment in real-world activity recognition applications.