Towards Edge Computing based Top-view Person Detection: YOLO۲۶ with Transfer Learning

Visual surveillance is essential for public safety in transportation hubs such as bus terminals and metro stations, where monitoring the crowd and detecting abnormal behaviors is crucial. Although advancements have been made over the years, especially with the use of top-view cameras to reduce occlusions and protect privacy, challenges remain regarding the vast amount of data generated by smart cities that must be processed in real-time. Edge computing addresses these limitations by bringing computation closer to the data source. In this work, a real-time person detection model is presented by utilizing the state-of-the-art YOLO۲۶ algorithm. The YOLO۲۶ model was originally trained for side-view person detection; therefore, we used transfer learning to improve the model's performance in top-view scenarios. Both models were trained under identical experimental conditions and evaluated by different detection metrics. Competitive analysis of the experimental results indicates that the fine-tuned model outperforms the pre-trained model by improving all metrics. The overall accuracy of the model is ۹۶%.

کلیدواژه ها:

Deep Learning ، Person Detection ، Edge Computing ، Transfer Learning ، YOLO