Modified YOLOv4 Framework with Thermal Images for Pedestrian Detection

A modified YOLOv4 framework is proposed for pedestrian detection which is trained and tested with a portion of ZUT-FIR-ADAS thermal image dataset. The effectiveness of mosaic and cutmix data augmentation techniques have been utilized to improve the generalization capability of this framework. The mean average precision of 73% and average loss of 1.8062 have been achieved through this framework with thermal images. It yields a precision of 0.92, recall of 0.96, F1 score of 0.94, average intersection of union of 72.25%. The experimental results show the effectiveness of the modified YOLOv4 framework for pedestrian detection under diverse weather conditions.