Fine-Tuning Faster R-CNN with ResNet50 for Infrared-Based Pedestrian Detection in Autonomous Vehicles: Performance and Comparative Analysis

Abstract

Hyperparameter tuning plays a critical role in optimizing deep learning models for pedestrian detection, particularly in challenging scenarios such as low-light and occluded environments. This study investigates the effect of fine-tuning key hyperparameters in Faster R-CNN with a ResNet50 backbone, focusing on learning rate, optimizer choice, batch size, weight decay, and scheduling. Two models were compared: a baseline Faster R-CNN and a fine-tuned version with optimized training strategies. The fine-tuned model incorporated a reduced learning rate (0.0001), AdamW optimizer with weight decay (0.0005), and a warm-up strategy to improve training stability. Trained for 50 epochs, the fine-tuned model demonstrated superior mean Average Precision (mAP@0.5) of 0.8505 compared to 0.816 in the baseline, with reduced fluctuations and improved convergence. These findings underscore the importance of hyperparameter optimization in enhancing detection accuracy and generalization, particularly for pedestrian detection.