Deep Learning-Based Consistent Object Detection in Distance on Highly Perspective Region

Abstract

CCTVs are commonly used for traffic monitoring and accident detection, but their images suffer from severe perspective distortion causing object size reduction with distance. This issue is exacerbated in tunnel CCTVs, positioned low due to space constraints, leading to challenging object detection, especially for distant small objects, due to perspective effects. To address this, this study proposes a solution involving a region of interest setup and an inverse perspective transformation technique. The transformed images, achieved through this technique, enlarge distant objects, maintaining object detection performance and appearance velocity across distances. To validate this, artificial CCTV images were generated in a virtual tunnel environment, creating original and transformed image datasets under identical conditions. Comparisons were made between the appearance velocity and object size of individual vehicles and for deep learning model performance with multiple moving vehicles. The evaluation was conducted across four distance intervals (50 m to 200 m) from the tunnel CCTV location. The results reveal that the model using original images experiences a significant decline in object detection performance beyond 100 m, while the transformed image-based model maintains a consistent performance up to the distance of 200 m.