Tactile Paving Detection and Tracking Using Tenji10K Dataset

Tactile paving is a ground-texture display device installed on sidewalks to guide visually impaired people. In recent years, many studies have been conducted on tactile paving detection using cameras to help visually impaired people walk independently. In addition, several computer vision-based tactile paving detection methods have been proposed. However, it is difficult to compare the detection accuracy of different proposed algorithms because there are no publicly available tactile paving datasets, and the ground truth and evaluation criteria used for evaluation are often inconsistent or not clearly presented. To this end, in this paper, we collect a tactile paving dataset and name it ‘Tenji10K’. The dataset is constructed with 20 sequences consisting of 10 K first-person tactile paving images taken in Japan, and ‘Tenji block’ refers to Japanese-style tactile paving. For detailed evaluation analysis, up to six sequence attributes are assigned, taking into account various real-world situations. On the other hand, we also proposed a tactile paving tracking algorithm based on an evolutionary algorithm. The effectiveness of the dataset is evaluated by conducting a comparative experiment based on Tenji10K with respect to four different methods, including our proposed method.