Computer Vision, Pattern Recognition & Deep Learning

A hierarchical fusion framework for robust humanoid activity recognition using IMU and RGB modalities, enhanced with genetic optimization for feature selection and Deep ConvLSTM for temporal modeling. Single-modality recognition collapses under occlusion and sensor drift — humanoids need redundant perception.

Conventional captioners name objects but not the relationships between them. This model reasons over structure before it writes a word using panoptic segmentation for pixel-level scene parsing and scene-graph decoding for relational understanding. The dual-branch architecture simultaneously processes visual features and textual semantics for contextually rich captions.

Treating participants as graph nodes lets the model recognize shared and interacting activities that per-person classifiers cannot see. The framework employs DeepConvLSTM for spatiotemporal feature extraction and graph neural networks for inter-person relationship modeling, enabling accurate recognition of group activities and collaborative tasks in multi-person environments.

An attention-driven framework for simultaneous disaster event classification and human pose estimation in emergency scenes using hierarchical feature fusion with deep ConvLSTM. The system processes multi-scale features through attention mechanisms for real-time identification of disaster types and victim poses.

A neuro-fuzzy classifier-based surveillance system for semantic segmentation driven crowd tracking and anomaly detection in crowded public spaces. The framework integrates pixel-level semantic understanding with fuzzy inference rules to track individual movements and detect anomalous behaviors such as running, fighting, or crowding.

Combines Graph Neural Networks with Gated Recurrent Units for collective sports activity modeling. The GraphSAGE architecture samples and aggregates neighborhood features while GRU layers model sequential dependencies, enabling accurate recognition of team formations, player roles, and tactical movements in team sports.

A wearable IoT framework integrating multimodal biosensors with intelligent signal processing for fall detection and remote health monitoring of elderly patients. The system fuses accelerometer, gyroscope, and heart rate data through deep learning models to distinguish between falls and daily activities.

Fusing pixel-level structural features from semantic segmentation with deep CNN appearance features for robust scene classification in robotics and autonomous navigation. The multimodal approach combines geometric layout information with visual texture patterns for accurate scene understanding across diverse environments including indoor spaces, urban landscapes, and natural terrains under varying lighting conditions.

Deep learning-based plant identification using leaf images for precision agriculture, weed management, and biodiversity monitoring. The framework employs convolutional neural networks with transfer learning to classify plant species from leaf morphology, vein patterns, and textural features.

HOG, LBP and Optical Flow hybrid features with multi-class SVM for vision-based human activity recognition from RGB cameras. The system extracts complementary descriptors capturing shape, texture, and motion information for recognizing activities such as walking, running, sitting, and gesturing.

Aerial surveillance system for real-time vehicle detection and tracking from UAV video using pixel-level semantic segmentation and particle filter tracking. The system segments vehicles from aerial views and maintains robust tracking through occlusions and camera motion, enabling traffic monitoring, disaster assessment, and military reconnaissance.