Automatic Crosswalk Detection and Counting Using YOLO and SAHI Techniques

The critical issue of pedestrian safety in large cities with high traffic density is focused on the effective detection and counting of pedestrians on crosswalks. This research proposes a deep learning model based on the You Only Look Once (YOLO) architecture, enhanced by the Slicing Aided Hyper Inference (SAHI) technique. The performance of several YOLO versions, including YOLOv8, YOLOv9, and YOLOv10, was evaluated with and without the integration of SAHI, using Accuracy, R-squared (R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>), and Mean Squared Error (MSE) as evaluation metrics. Our experimental results represent that, without SAHI, YOLOv9 achieved the highest accuracy of 37.52%, with an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.98 and an MSE of 19.36. When the SAHI technique was applied, YOLOv9+SAHI demonstrated a significant improvement by achieving an accuracy of 75.12%, maintaining an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.98, and reducing the MSE to 11.67. These findings highlight the effectiveness of SAHI in enhancing detection performance and reducing errors, particularly for YOLOv9, and improving the detection of small objects in complex environments. This study provides valuable insights into optimizing pedestrian detection systems, with YOLOv9 integrated with SAHI being the most effective model for pedestrian counting on crosswalks.