For infrared vision or multi-spectral vision pedestrian detection research, some researchers have constructed some data sets for related applications in recent years, among which the following five are more representative.

1) KAIST Multispectral Pedestrian Detection Benchmark

Hwang et al. built an acquisition system consisting of a visible light camera, an infrared thermal imager, a spectroscope, and a three-axis camera fixture to achieve precise optical registration images, using this system to shoot daytime and nighttime traffic scenes in campuses, streets, and villages. Construct KAIST (https://github.com/SoonminHwang/rgbt-ped-detection), a visible and infrared image pedestrian detection dataset for intelligent driving. Each image in KAIST contains two versions of RGB color image and thermal infrared image, the image resolution is 640x480, a total of 95,328 images. The labels of the data set include three categories of person, people, and cyclist. It contains a total of 103,128 pedestrian annotation boxes. Among them, the individuals that are better distinguished are labeled as person, and the individuals that are not easy to distinguish are labeled as people. Labeled as cyclist. Among all the marked objects, the unoccluded, partially occluded and severely occluded pedestrian targets accounted for 78.6%, 12.6% and 8.8% respectively.

 

2) OTCBVS Benchmark Dataset Collection

OTCBVS Benchmark Dataset Collection (http://vcipl-okstate.org/pbvs/bench/) is a multi-task-oriented image database consisting of 14 sub-databases. Among them, the Ohio State University Body Temperature Database (OSU Thermal Pedestrian Database), Terravic Motion IR Database (Terravic Motion IR Database), Visual Analysis Thermal Infrared Video Benchmark for Visual Analysis, etc. analyze. Taking the thermal infrared video benchmark library for visual analysis as an example, it includes thermal infrared sequence images with multiple resolutions from 512x512 to 1024x1024. applications such as counting.

 

3) FLIR Thermal Dataset

FLIR Thermal Dataset (https://www.flir.in/oem/adas/adas-dataset form/) is FLIR System's ADAS application environment, for challenging weather such as total darkness, smoke, severe weather and glare. A publicly available dataset of thermal infrared images developed with machine learning algorithms (especially DNN algorithms) that can better detect and distinguish pedestrians, cyclists, animals, and motor vehicles under certain conditions. The data set has a total of 14,000 images, describing the scenes of driving on the streets and highways of Santa Barbara, California in sunny to cloudy weather (60%) and night (40%), and the objects in these scenes include people, cars, etc. , bicycles, dogs, and other vehicles are marked accordingly.

 

4) SCUT FIR Pedestrian Dataset

SCUT FIR Pedestrian Datasets is a large-scale long-wave infrared pedestrian detection dataset (https://github.com/SCUT-CV/SCUT_FIR_Pedestrian_Dataset), which consists of 11 hours of sequence images with a frame rate of 25 Hz. Under the speed of 80 yards per hour (that is, 80 km/h), traffic scenes of 11 road sections in Guangzhou's commercial districts, suburbs, expressways, and campuses. The dataset has a total of 477,907 labeled boxes for 7,659 different pedestrians in 211,011 frames. In this dataset, SO-S10 is the training set, with a total of 70,517 frames, 3,703 different pedestrians, and 240,297 labeled boxes; S11-S20 is the test set, with a total of 73,115 frames, 3,940 different pedestrians, and 237,610 labeled boxes.

 

5) INO Videos Analytics Dataset

INO Videos Analytics Dataset (ftp://ftp.ino.ca/VideoAnalyticsDataset) was constructed by the National Optics Institute of Canada (National Optics Institute of Canada), including a variety of outdoor scenes under different weather conditions such as parking lots, campuses, highways , the visible light sequence images of the hall entrance, etc., and the visible light-thermal infrared multispectral sequence images. These images include single/multi-person behavior, vehicle movement, etc. In addition, the thermal infrared and visible light images have achieved pixel-level registration.