Abstract:
Human detection plays a pivotal role in many vision-based applications. Effectively detecting humans across diverse
environments and situations significantly contributes to enhancing human safety. However, this effectiveness encounters challenges,
particularly in hazy conditions that reduce visibility and blur images, thereby impacting the accuracy of existing detection algorithms.
Additionally, the quality of dataset annotations significantly affects the accuracy of these systems. Poor annotations lead to insufficient
training of detection models, resulting in higher error rates and reduced efficacy in real-world scenarios. To tackle these challenges,
we’ve introduced new, more precise annotations for the INRIA dataset. These enhancements overcome limitations within the dataset,
particularly instances where numerous individuals in images lacked proper labeling. This augmentation aims to improve training
robust detection models and provide a more accurate evaluation of the model’s performance. Our experiments have yielded notable
improvements, showcasing a 20.37% increase in Average Precision and a substantial 68.19% reduction in False Negatives. Moreover,
we’ve developed a deep-learning model for human detection, leveraging transfer learning to fine-tune the YOLOv4 model. Experimental
results demonstrate that our proposed model accurately detects pedestrians under various weather conditions, including both clear and
hazy scenarios. It achieves high average precision and F-Scores while maintaining efficient real-time operation at 55.4 FPS. These
advancements significantly enhance the reliability and applicability of human detection systems.