Accurate Pedestrian Detection for Human Crowds Using Deep Learning Techniques رسالة الماجستير

Abstract

Pedestrian detection, especially in crowded environments, is a pivotal task for various applications such as surveillance, autonomous vehicles, and crowd management. The accuracy and efficiency of pedestrian detection systems are highly contingent on the methodologies employed. This thesis investigates and evaluates the development of deep learning-based pedestrian detection models, with a particular focus on anchor box optimization and the integration of clustering algorithms. Two distinct approaches are employed for anchor box optimization: K-Means and Fuzzy C-Means clustering. These methods aim to fine-tune anchor boxes to be more representative of the dimensions and shapes of pedestrians in the dataset. An in-depth empirical analysis is conducted to evaluate the performance of the two models. The model incorporating K-Means for anchor calculation achieves a mean Average Precision (mAP) of 87.6% and an F1-score of 83.7%. In contrast, the model utilizing Fuzzy C-Means achieves a slightly higher mAP of 88.1% and an F1-score of 84.2%. In the inference results section, real-world images without prior annotations are used to evaluate the practical performance of both models. Through visual inspection and comparison of bounding boxes, the study ascertains the effectiveness of each model in detecting pedestrians under real-life conditions. The thesis then concludes by highlighting the impact of anchor box optimization through clustering algorithms and providing insights into the practical deployment of deep learning models for pedestrian detection in crowded environments. In the initial phase, YOLOv7, a stateof- the-art object detection model known for its precision and speed, is adapted for V pedestrian detection through transfer learning on the CrowdHuman dataset. This research adds to the field of pedestrian detection by underlining the significance of anchor box optimization using clustering algorithms. Importantly, the role of YOLOv7 emerges as a pivotal factor influencing our results. By integrating YOLOv7, we establish a robust baseline that informs our exploration of anchor box optimization and clustering algorithms, contributing to a comprehensive understanding of their impact in crowded environments