Sense as you go: A context-aware adaptive sensing framework for on-road driver profiling

• The realization of next-generation intelligent transportation services primarily depends on the rich data that could effectively capture the on-road and in-vehicle conditions in real time. Road safety being one of the most critical objectives for such services, received significant research interest in the past. Smartphone-based inertial sensing has proved to be quite promising in identifying potential anomalies in road conditions or driving patterns. However, the data generated from the onboard Inertial Measurement Unit are exposed to high noise due to mobility and subjective diversity. Moreover, a large amount of data produced at a higher sampling rate directly impacts on the battery life of such resource-constrained devices as well as the privacy of the concerned user. To address this set of challenges, we introduce a novel framework for mobile devices with the objective of supporting ‘sensing on a need’ basis. The proposed framework, AdaSift, has three key features: (1) detects the relevant context signifying a window of important data points in real-time, (2) filters out irrelevant data at the source itself and thus reduces the associated overhead significantly; and (3) effectively identify a driver’s profile based on the unique behavioral signature traced by the filtered data. We evaluated the performance of AdaSift by implementing the framework on an Android smartphone and conducting a user study with 8 drivers. The experimental results show that the use of AdaSift could achieve up to accuracy in on-road driver profiling while reducing the energy consumption footprint of the smartphone by up to . 

 FedMeet: Personalized Federated Multi-Sensor Fusion for Privacy-Preserving Human Activity Recognition

• Federated Learning (FL) offers a promising avenue for privacy-preserving Human Activity Recognition (HAR) using distributed on-body sensor data. However, existing FL approaches often underperform in real-world deployments due to client heterogeneity, modality-specific sensor dynamics, and the non-IID nature of human behaviors across users. We present FedMeet, a personalized federated learning framework for human activity recognition (HAR) using multi-sensor on-body sensing data from smartwatches and earables. FedMeet addresses key challenges in distributed HAR, including non-iid client data, sensor heterogeneity, and high intra-class variance with low inter-class variance, common in real-world deployments. Unlike existing methods, FedMeet introduces a unified architecture that integrates sensor fusion, temporal modeling, personalization, and privacy preservation. These gains stem from four key components: (1) gated sensor fusion for adaptive modality relevance, (2) a BiLSTM backbone to capture temporal activity patterns, (3) client-level personalization to accommodate behavioral and sensor placement heterogeneity, and (4) a privacy-preserving masking mechanism that securely perturbs client weight updates before transmission, reducing the risk of information leakage. Experimental results on a real-world multi-sensor HAR dataset show that FedMeet achieves a test accuracy of 87.97% within 20 rounds, outperforming state-of-the-art baselines (FedPer, FedProx, and ClusterFL) across precision, recall, and F1-score. FedMeet offers a promising solution for privacy-aware, real-time HAR in collaborative environments such as workplaces and smart meeting spaces. 

Experience: From Sensors to Service: Advancing Ambient Living with an AIoT Community Testbed

• Rapid advancements in IoT and AI over the past decade, coupled with their pervasive integration into daily life, have created substantial opportunities for research and development in Artificial Intelligence of Things (AIoT) toward the comprehensive realization of smart city services. These human-centered applications require suitable testbeds for in-the-wild usability studies and performance validation, ensuring reliable sensory data for training AI models. Yet, most existing testbeds remain either application-focused (e.g., occupancy, energy, mobility) or scenario-bound, which limits their extensibility, heterogeneity, and usability for researchers. This paper introduces SensePod, an AIoT-based smart home testbed established in the IIT Jodhpur campus, designed to promote Sensing as a Service for ambient living. Unlike city-scale or lab-restricted platforms, SensePod transforms a one-bedroom residential space into a modular and reconfigurable substrate where environmental, motion, acoustic, and wearable streams can be orchestrated and reused across studies. We document the requirement analysis, system design, and operational lessons learned in deploying SensePod, and demonstrate two representative use cases: privacy-preserving occupancy estimation and multimodal activity recognition via wearable integration. Our results show that multimodal fusion achieves over 90\% accuracy for activity inference, while edge-based deployment enables low-latency, privacy-aware operation. By exposing real-world challenges such as calibration drift, heterogeneous sampling rates, and privacy trade-offs, SensePod provides a researcher-oriented platform that bridges the gap between controlled lab testbeds and the unpredictability of lived environments, paving the way for community-driven innovation in ambient intelligence.

Fine-grained Indoor Air Quality monitoring: Development, Calibration, and Prediction under the project IMPRINT , Under the Guidance of Prof. Subrata Nandi and Dr. Sujoy Saha,Department of CSE , NIT Durgapur [Jun 2018 - Oct 2020]

Road Surface Event detection using heterogeneous sensors under the project of CityProbe ,NIT Durgapur [Jun 2018 - Dec 2018] 

Human Activity Reorganization using ultrasonic Sensor, PIR Sensor under the project DISARM, Department of CSE and CA, NIT Durgapur Funded by Media Lab Asia, Govt. Of India [ Dec 2015-June 2016].

Classification of Aromatic rice samples using FT-NIR spectrometer and MATLAB (s/w) from C-DAC Kolkata. [Jun 2015 - Jul 2015]