March 24th, 2018
by May Thu
Introduction
In Asia, the rate of road accident is much higher every year and nearly 1.3 million people lost their life on road. In 2018, Thailand’s road accident is at the peak of ranking among other countries and pedestrian accidents still represent the second largest source of traffic related injuries and accidents after accidents. Nowadays, pedestrian detection system which is able to detect the human by recognizing the human structures or its model appearing in image has much gained attention of many researchers in computer vision. There are many complex pedestrian movements and various automobiles traffic in Southern Thailand. In this dataset, we introduce crowded pedestrians between different vehicles and appearance variations in real world environment which is much different with the standard Europe dataset. This dataset is to provide the occlusion, partial occlusion, appearance, and posture of pedestrians with real world data and to help the performance of algorithm improvement.
There are two types of dataset among the existing datasets which are classification datasets and detection datasets. PSU dataset contain this two types: cropped pedestrians and original pedestrian images for training and testing the detection system.
We created two samples: positive which contain human and negative which do not contain human.
Figure1. Examples of positive samples
Figure 2. Examples of Negative samples
Dataset Properties
We captured pedestrian data from multi-view positions according to various posture: upright, walking, standing, cycling, motorbike riding, left, right, back, and occluded part of human at different viewpoints with OPPO A57 and Samsung Note5 around the marketplace areas, PSU campus and some data are taken from google. Original image resolution is 3120×3120 pixel and there are four types of pixels which is resize with 64×64, 256×256, 720×960 and 960×720 pixel resolution per images for training and testing data. Our dataset contains different resolutions to get effective for evaluating the detection algorithm.
The useful properties: scene, view, pose, occlusion, and attachment information are provided for each pedestrian. We captured the images at different crowded areas in marketplace and university campus with complex background. The scenes of the image contain pedestrian crossing road, marketplace areas, city centre junction points, and campus roads to represent the real world environment. For illumination, we separated the daytime and evening with different light condition. In Southern Thailand, there are only two seasons: sunny and raining. Therefore, the recorded period is in summer but the images are collected into different illumination: normal, cloudy, sunny, and night in daily life.
The scale of the images is near, medium and far due to the clearly captured the pedestrian height. The distance between camera and pedestrian are wide range and some are narrow range especially inside the market. The scale of the images also important to evaluate the performance of the system. Therefore, the dataset provides different scale of the images with the various height of the pedestrian. Various position of pedestrians is also captured from different angles of view
For training, the original images are resized with different resolution and are cropped into different types of human posture (Single, Two Person, Three Person, and Crowded). We split the dataset into training and testing sets which are centred and cropped human images for positive training set and images with no human for negative training set. For the testing images, we created different pixels of images containing the original resolution images for testing and evaluating the performance.
The following relevant directories contain in this dataset:
Positive Samples: 64×64 pixel resolution (1051 samples)
256×256 pixel resolution (1051 samples)
720×960 pixel resolution (1051 samples)
960×720 pixel resolution (1051 samples)
Negative Samples: 64×64 pixel resolution (517 samples)
256×256 pixel resolution (517 samples)
720×960 pixel resolution (517 samples)
960×720 pixel resolution (517 samples)
For training, the original images are resized with different resolution and are cropped into different types of human posture (Single, Two Person, Three Person, and Crowded).
We split the dataset into training and testing sets which are centred and cropped human images for positive training set and images with no human for negative training set.
Train Positive Samples: 64×64 pixel resolution (1084 samples)
256×256 pixel resolution (1084 samples)
720×960 pixel resolution (1084 samples)
960×720 pixel resolution (1084 samples)
Train Negative Samples: 64×64 pixel resolution (517 samples)
256×256 pixel resolution (517 samples)
720×960 pixel resolution (500 samples)
960×720 pixel resolute (517 samples)
Test data samples: 1272 samples (positive and negative)
Figure 3. Examples of positive training cropped samples
Download
Please read and accept Agreement and Disclaimer below before downlaod. If you downlad this dataset, we count your accepted agreenent and disclaimler.
You may downlad the whole dataset form here.
For trianing, you may downlad here.
Disclaimer
- This dataset must be only used for non-commercial or educational purpose.
- This data set is provided “as is” and without any express or implied warranties, including, without limitation, the implied warranties of merchantability and fitness for a particular purpose.
The images provided above may have certain copyright issues. We take no guarantees or responsibilities, whatsoever, arising out of any copyright issue. Use at your own risk.