新鲜出炉的盲道检测开源数据集（TP-Dataset）相关介绍

基于视觉分割（Visual segmentation）技术进行盲道检测，方便快捷，为盲人出行提供便利!

盲道作为盲人出行的主要辅助设施，在城市街道中不可或缺，但拥挤的人流和残缺的盲道等影响了盲人出行的便利，促使盲道自主检测技术的研发成为了社会公益中一项亟需解决的任务之一。随着视觉分割技术的发展，使得各类典型目标的分割定位任务顺利完成。TP-Dataset基于视觉分割任务特点，将盲道和周边背景信息进行二分类标注，结合分割模型进行盲道自动识别任务，有一定的借鉴意义！该盲道数据集及相关论文和开源代码介绍如下：

Tactile paving plays a crucial role in the travel of visually impaired people. Therefore, it is quite meaningful to identity the shape and location of tactile paving to support the mobility of the visually impaired, and visual segmentation technology is suitable of this task.

The new-built dataset is focus on tactile paving based on the technology of visual image segmentation, which is called tactile paving dataset (TP-Dataset).

We hope that the visually impaired people could get more attention and full protection in daily life, and the researches on this topic may provide more assistance and support for them!

Contributors (listed in no particular order, and sorted by phonetic initials of last names):

Jing Lee; Mingshuai Li; Ruize Li; Lei Liang; Jinyu Liu; Yaru Shang; Wenzhe Tang; Gusheng Tian

Kunli Wu; Zhihui Wang; Xingli Zhang; Shenglu Zhao

Above-mentioned Contributors have participated in the process of data collection and annotation of the TP-dataset, and we thanks for their efforts.

Introductory Paper:

@article{ZHANG2024122109,

title = {GRFB-UNet: A new multi-scale attention network with group receptive field block for tactile paving segmentation},

journal = {Expert Systems with Applications},

volume = {238},

pages = {122109},

year = {2024},

issn = {0957-4174},

doi = {https://doi.org/10.1016/j.eswa.2023.122109},

url = {https://www.sciencedirect.com/science/article/pii/S0957417423026118},

author = {Xingli Zhang and Lei Liang and Shenglu Zhao and Zhihui Wang}

}

Download link:

(1) GRFB-UNet code:

GitHub - Chon2020/GRFB-Unet: GRFB-UNet: A new multi-scale attention network with group receptive field block for tactile paving segmentation

(2) TP-Dataset:

Baidu Net-disk:

百度网盘 请输入提取码 (password: 9ope)

Google Drive:

https://drive.google.com/drive/folders/1jByE5f_oUKpYdoR829wLqFSqlBYtOZM6?usp=sharing

Characteristics of the TP-Dataset:

With comprehensive considerations of various factors, this dataset is built with following characteristics:

 (1) Scenario: Multiple typical scenarios are included in TP-Dataset, such as campus, street, railway station, bus stations, subway, community and hospital. The total amount of images with tactile paving is 1391.

(2) Data source: The majority of these images are captured in actual situations by different types of camcorders or phone cameras. Remaining images are searched and downloaded online. The quantities of captured images and online images are 828 and 563 respectively.

(3) Observing angle: For the captured images, the camcorders or phone cameras are fixed on camera stand or hand-held with horizontal angle or slight declination, and the heights are between 0.8m and 1.8m above the ground. Most online pictures are also searched with this rule, and others are collected by surveillance camera with looking down angles and higher positions.

(4) Appearance style: The styles of tactile paving contains strip-type bricks for the guidance of the direction of travel, and dotted bricks for the warning of obstacles or corners.

(5) Image resolution: Since the image resolution is varied with many times, the images in TP-Dataset is scaled with following rules: For each image, the minimum edge between image height and width is selected and resized to be 283, with fixed image aspect ratio.

(6) Illumination condition: Most of the images are collected in daytime, and partial images are captured at night or in evening under streetlight or with camera’s flash. The amounts of images in daytime and at night are 1289 and 102 respectively.

(7) Data annotation: The regions of tactile paving on images are annotated with pixel-based masks. Since tactile paving is only distinguished from background, the target is saved in the format of gray image, with the value 255 for tactile paving and 0 for background.

(8) Data partition: These images of tactile paving are stored with nine parts, based on the types of scenario and data source. Images in TP-Dataset were randomly assigned into training set / validation set / test set, with the quantity of 824 / 281 / 286 respectively

With aforementioned characteristics, the TP-Dataset contains various kinds of external conditions, which contains the majority of actual situations. Therefore, this dataset is representative and useful for the training of image segmenting models.

Attribute Information:

Based on the data source (captured or downloaded online), scenario type and illumination condition, this TP-Dataset are roughly divided into different parts, which are listed as follows:

Part01: Campus (captured at daytime)

Part02: Street (captured at daytime)

Part03: Station (captured on railway station, bus stations and subway at daytime)

Part04: Community (captured at daytime)

Part05: Hospital (captured at daytime)

Part06: Community or campus or street (captured in evening or at night)

Part07: Community or campus (online),

Part08: Streets (online)

Part09: Station (downloaded online on railway station, bus stations and subway)

TP-Dataset is prepared in the following structure:

                                                         TP-Dataset

                                                               |——————JPEGImages

                                                               |        └——————Part01

                                                               |        └——————Part02

                                                               |        └——————… …

                                                               |——————Ground truth

                                                               |        └——————Part01

                                                               |        └——————Part02

                                                               |        └——————… …

                                                               |——————SegmentationClassPNG

                                                               |        └——————Part01

                                                               |        └——————Part02

                                                               |        └——————… …

                                                              └——————Index

                                                                         └——————train.txt

                                                                         └——————val.txt

                                                                         └——————predict.txt

In TP-Dataset, the images are stored in the file ./JPEGImages, and ground truths are stored in file ./ground truth. Masked images are stored in the file ./ SegmentationClassPNG and only used for visualization, which can be removed without any influence. Indexes of train / validate / test samples are stored in flie ./Index.

Note that we just split these samples in nine parts, and their labels are not distinguished. The annotated ground truths are divided into the regions of tactile paving or background.

Experimental results:

Comparison of experimental prediction results of advanced network models:

Method	MIoU	mPA	Accuracy	IoU	Recall	Precision
PSPNet	93.40	95.75	98.51	88.49	92.08	95.77
SegNet	82.91	90.91	95.64	70.69	84.62	81.12
FCN	72.79	78.21	93.51	52.57	57.84	85.21
Simt	85.34	87.77	96.74	74.28	75.83	97.32
DeepLabV3	92.60	95.23	98.32	87.09	91.11	95.18
DeepLabV3+	91.53	94.85	98.85	85.26	90.59	93.55
Mask R-CNN	83.52	88.00	96.14	77.31	77.17	90.38
GRFB-UNet	94.85	96.76	98.85	91.21	94.20	96.65

 Notation: The results in the table are all percentage data (%).

License:

This dataset is published under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License. This means that you must attribute the work in the manner specified by the authors, you may not use this work for commercial purposes and if you alter, transform, or build upon this work, you may distribute the resulting work only under the same license.