Efficient Position Estimation Based on GPU-Accelerated Content-based Image Retrieval

Abstract

We propose an efficient position estimation method based on GPUaccelerated content-based image retrieval (CBIR). The idea is to use videos of first-person vision associated with geographical position information as the database. When a user sends a current subjective image, the system estimates the position using CBIR. Since features extracted from images are in general high-dimensional vectors, thousands of vectors are extracted even from a single image, resulting in high processing cost. On the other hand, GPUs (graphics processing unit) have contributed to accelerate various processing, while they are originally for graphics processing. Therefore, we utilize GPU to accelerate CBIR with appropriate data structures and algorithms. Moreover, our proposed method considers spatial locality of pedestrians in position estimation applications in order to improve accuracy. We demonstrate the efficiency and accuracy of the proposed method through experiments using a video dataset.