Analysis of Error Trends by Range in Visual Localization: Comparing Image Retrieval and Image to Point Cloud Registration

초록

Localization technologies based on radio signals, such as the Global Navigation Satellite System(GNSS), show varying results depending on communication conditions. However, robust localization requires the ability to function in these challenging conditions. Visual localization, which operates independently of communication environments, can complement existing systems. Despite the diverse characteristics of visual localization depending on the sensors and methods used, there is a lack of research analyzing these differences. This paper compares and analyzes two visual localization methods: image retrieval and image to point cloud registration(I2P). Evaluation metrics include dataset size, processing speed, and translation error. In particular, translation error is analyzed by specific ranges to examine the trends associated with different errors. This detailed analysis helps identify the unique characteristics of each method. For a fair evaluation, research from the same year is selected for each method. The evaluation results show that the average translation error for image retrieval is 15.428 meters, while for I2P, it is 28.519 meters, indicating that image retrieval is more accurate. However, for errors below 0.5 meters, image retrieval shows an error of 0.339 meters, whereas I2P demonstrates greater precision with an error of 0.160 meters.