Aditya Shankar Vishnubhatla

Contributing to the development of a real-time aerial object detection and tracking system for UAVs using deep learning and computer vision. Focused on enhancing model precision and speed for aerospace-grade situational awareness and threat detection application

Key Contributions

Engineered custom CNN-based detection models (YOLOv8 + EfficientNet backbone, yolo12, 13 customised versions as well ) that improved mean Average Precision (mAP) by 15.2 %, reduced false-positive rates by approx 30 %, and increased F1-score to 0.93 on complex aerial datasets.

Automated data-annotation and augmentation workflows using Python scripting and CVAT Roboflow, shrinking dataset-preparation time and work by ≈ 45 % (no of images reduced) and enabling continuous integration of new imagery from UAV sorties.

Built a fully reproducible end-to-end ML pipeline covering data ingestion, preprocessing, model training, validation, versioning, and CI/CD-based deployment; improved training throughput by 34 % using mixed-precision and distributed learning.

Implemented TensorRT and ONNX-Runtime inference optimization, achieving a 51 % reduction in latency (from 70 ms → 34 ms per frame) and 38 % lower GPU utilization during real-time edge inference personalised weights and hyoerparameter tuning and by usage of cpu and T4GPU

Designed Kalman-filter + IoU-based multi-object tracking, maintaining stable target locks at > 63 FPS and
Applied hyperparameter optimization, early-stopping, and k-fold validation to maximize model generalization and prevent overfitting across varying altitude and illumination conditions.

Will be Deploying the optimized pipeline on Jetson AGX Orin hardware for live UAV field tests, validating operational robustness at altitudes up to 3 km with > 91 % detection accuracy.

Collaborated cross-functionally with aerospace and embedded-systems teams to ensure AI reliability, explainability, and safety compliance under CSIR’s autonomous-flight guidelines.

Authored a technical report and internal deployment documentation contributing to CSIR–NAL’s ongoing research in AI-enabled perception for autonomous aerial vehicles