Robust method for variation in pedestrian appearance from fisheye videos

CONTEXT This internship is part of the development of an autonomous drone equipped with a fisheye camera for intelligent surveillance. Pedestrians present significant appearance variations (clothing, accessories, posture, lighting) that degrade the performance of classic detectors. The goal is to develop a robust method to handle these appearance changes in fisheye images. INTERNSHIP OBJECTIVES • Robust pedestrian detection despite appearance variations • Reliable re-identification in fisheye video sequences • Integration with a real-time detector (YOLO) TASKS REQUIRED • Bibliographic study on robustness to appearance variations • Development of a spatio-temporal attention approach • Python/PyTorch implementation • Integration with YOLO for real-time detection • Comparison with the state of the art • Writing of a scientific article DELIVERABLES • Method robust to appearance variations • Validated integration with YOLO • Comparative benchmark with the state of the art • Scientific article