Life-threatening Ventricular Arrhythmias (VAs) Detection (TinyML contest in ICCAD 2023) and VLSID 2023

Developed and deployed a CNN on a STM32F303K8 board to detect VAs. The CNN analyzes IEGM segments as input data and provides real-time classification of VAs. Achieved an F-Beta of 93.3% with an average response time of 8.4ms Secured overall Rank 13, Team Name: IIT-Goa-Systems-Lab

Joint-YODNet: A Light-weight Object Detector for UAVs, CVIP 2023 

Small object detection via UAV (Unmanned Aerial Vehicle) images captured from drones and radar is a complex task. we proposed a novel method called Joint-YODNet for UAVs to detect small objects, leveraging a joint loss function specifically designed for this task. Joint-YODNet achieves a recall of 97.1% and an F1 Score of 97.5%, surpassing state-of-the-art (SOTA) techniques.

Real-time Video Processing System For Person-Tracking Across Multiple Cameras

 The growing need for video surveillance in public spaces has created a demand for systems that can track individuals across multiple cameras feeds in real time. We proposed a person-tracking system that combines correlation filters and Intersection Over Union (IOU) constraints for robust tracking, along with a deep learning model for cross-camera person re-identification.

YOLOv9++ for Plant Biometrics

 Agriculture remains crucial to many economies worldwide, yet plant diseases threaten its productivity and sustainability. Early detection and timely treatment are essential for mitigating their impact. To address this, we propose YOLOv9++, an enhanced variant of YOLOv9 integrated with Convolutional Block Attention Module (CBAM). Our model significantly improves accuracy and efficiency in plant leaf disease detection. We introduce a large-scale leaf disease detection dataset called PlantVillage-LD2 inspired by the PlantVillage dataset. YOLOv9++ achieves competitive performance, surpassing the baseline YOLOv9, achieving a mAP:50 of 93.7% on the PlantVillage detection dataset.

Semi-Supervised Drone Technology for Real-time Traffic Rule Enforcement

In today's transportation landscape, two-wheelers have become the dominant choice despite their inherent risks due to minimal protection. Though governments have imposed penalties for non-compliance, traditional enforcement methods suffer from limitations like manual surveillance, limited coverage, and high costs. To address these challenges, we propose a novel aerial view solution for automating helmet detection to enforce regulations effectively. We introduce SkyGuard, a drone-based multi-head YOLOv5 with a transformer block (C3TR) for improved detection. We surpass all the standard state-of-the-art detection models with a mAP:50, precision and recall rates of 98.5%, 96.6% and 94.7% on the OSF Helmet Dataset, respectively.

Aerial Tree Crown Delineation and Species Classification

Traditional forest surveys have long relied on labor-intensive on-foot assessments conducted by specialized teams. In this work, we introduce a novel Attention Block called Hybrid Pooling Efficient Channel Attention (HyP-ECA), which is integrated into the architecture of YOLOv8, resulting in HyP-ECA-YOLO. We also address the lack of aerial datasets for tree species detection, we curated the Multi-Tree species aerial detection (MTAD) dataset by consolidating and re-annotating open-source aerial datasets covering five types of trees. HyP-ECA YOLO achieves a mean Average Precision (mAP:50) of 87.9% on our dataset. Furthermore, We validated our Aerial Data collection Survey process through a virtual Software In The Loop (SITL) simulation environment using ROS and Gazebo.