Carnegie Mellon University People Image Analysis Consortium Technical Report Image Dataset for Researches about Surveillance Camera - CMU_SRD (Surveillance Research Dataset) - Koosuke Hattori Hironori Hattori Yuji Ono Katsuaki Nishino Masaya Itoh Vishnu Naresh Boddeti Takeo Kanade November, 2014 (Ver.1.0.0)
Contents 1. INTRODUCTION... 3 2. ABSTRACT OF THE DATASET... 3 3. HARDWARE CONSTRUCTION... 4 CAMERA... 4 LENS... 5 STORAGE... 5 4. SCENARIO FOR DATA COLLECTION... 8 LOCATION AND CAMERA POSITIONS... 8 COURSE FOR DATA COLLECTION... 10 KINDS OF THE DATASETS... 12 5. COMPARISON WITH OTHER DATASET... 13 6. CONCLUSION... 15 7. ACKNOWLEDGMENT... 15 8. REFERENCES... 15 2
1. Introduction Many surveillance cameras have been installed into public spaces. They became popular and necessary tools for safety life. Some of the surveillance cameras have functions that were applied computer vision techniques, such as object detections, tracking, detecting events and so on. To survey the detail of the objects requires high- resolution images. High- resolution images make possible to survey the detail with eyes. Increased resolution will also give additional possibility to computer vision field. We start this project to collect a comprehensive new dataset for indoor surveillance scenario. The collection site is a CMU building where multiple cameras are installed. More than several tens of people are used as subjects. We named this dataset CMU_SRD (Surveillance Research Dataset). Everyone can use this dataset for his/her research purpose. This dataset is expected to become a standard research tool. 2. Abstract of the Dataset CMU_SRD is sequence images including pedestrians. These images were captured with 8 cameras. Two of the cameras are arranged as a stereo camera pair. Pedestrians walked along the predetermined route. We will describe the route in chapter 4. CMU_SRD has two kinds of images. The first kind images were captured with pedestrians walking individually. These images were named Individual dataset. The other dataset ( Group ) were captured with crowded pedestrians (3 or 4 participants). "Group" images have more occluded area than "Individual" ones and will be more difficult for detection/tracking/re- identification pedestrians. We provide images for camera calibration. Users can calculate camera parameters for each camera by themselves. We also provide synchronized images for stereo camera pair calibration. 3
3. Hardware Construction Camera We selected Point Grey Flea3 (FL3- FW- 14S3C- C) as cameras to use the data collection. Cameras specification is shown in Table 1. We selected IEEE 1394b model, because IEEE 1394 model cameras can be easily synchronized by using IEEE 1394 bus signal. The reason why we selected color model is to use this dataset for the researches that require color information (e.g. People Re- Identification, Color Based Tracking etc.). We did not use auto exposure, gain, white balance control and any automatically controlled parameter. All of the parameters were same for all cameras. Table 1. Specification of the camera. [1] Appearance of the Camera Manufacturer Model Sensor Shutter Maximum Resolution Point Grey Research Inc. Flea3 (FL3- FW- 14S3C- C) Sony ICX267 CCD, 1/2", 4.65 µm Bayer Color Global Shutter 1384x1032 at 16 FPS We used MultiSync [2] that is provided by Point Grey Research when we collected data. MultiSync is a software to synchronize the image acquisition of multiple PGR cameras using IEEE 1394 bus. 4
Lens We selected Fujinon DF6HA- 1B Lens for the data collection, because this lens is designed for high- resolution (up to 1.5 mega pixel) cameras. The specification of this lens is shown in Table 2. Table 2. Specification of the lens. [3] Appearance of the Lens Manufacturer Model Mount Focus Length Fujifilm Co. Fujinon DF6HA- 1B C 6 mm F Number 1.2 to 16 Field of View 56 09ʹ 43 36ʹ with 1/2" Sensor Storage Writing speed of the storage is important to collect images without dropped frame. We tried to maximize the writing speed of the storage in the recording machine. We used RAID0 disk which is consisted with 8 SSDs. We selected HighPoint RocketRAID 2720 and Corsair Force GT 240. The details of the components are shown in Table 3 and 4. 5
Table 3. Specification of the RAID card. [4] Appearance of the Card Manufacturer HighPoint Technologies, Inc. Model RocketRAID 2720 Data Transfer Speed Interface Up to 6Gb/s PCI- Express 2.0 x8 Number of Device Channels 8 Compatible Disk Device SAS/SATA 3Gb/s Table 4. Specification of the Solid State Drive. [5] Appearance of the Card Manufacturer Model Interface Speed Corsair Force GT 240 GB SATA 3 6Gb/s Max Random Write Speed : 85,000 IOPS Max Read Speeds : 555 MB/s Max Write Speed : 525 MB/s 6
The benchmark result of this storage is show as Fig. 1. Sequential writing speed was around 1.6 GB/s. This speed is around 6 times comparing with a single Force GT SSD s benchmark result (about 270 MB/s). Fig. 1. Benchmark Result of the Storage Speed 7
4. Scenario for Data Collection Location and Camera Positions We recorded the data on the 4th floor of Carnegie Mellon University Newell Simon Hall. The image coverage map of the site is shown as Fig. 2. Cameras were set like real surveillance cameras. All of the entrance to this area are covered by the cameras. Camera 7 and Camera 8 are arranged as a stereo camera pair. User can compute and use depth with this stereo camera. Camera1 Camera7 Camera8 Camera6 Camera2 Entrance Camera Covered Area Camera3 Camera4 Coffee Space Go to QoLT Camera5 Main Entrance Fig. 2. Coverage Map and Example Images. 8
Lighting conditions are changed dramatically especially by position in the images captured with Camera 4. Illuminations installed in the low ceiling caused this change. Fig. 3. Shows the measurement results of the site. 2.00 m 8.88 m 1.80 m 1.87 m 0.61 m 8.64 m 20.5 m 3.94 m 2.58 m 3.00 m 2.64 m 2.00 m 16.4 m 4.00 m 0.64 m 4.40 m Coffee Space 3.54 m Entrance Camera 13.40 m Main Entrance Camera4 Go to QoLT Fig. 3. Measurement Results of the Site. 9
Course for Data Collection We predetermined the course for pedestrians. Fig. 4 shows the course we designed. The course was consciously designed to capture various angles of pedestrian images. The walking course was sometimes changed, because the coffee space was under renovation during the data collection. Fig. 5 shows the examples of images captured with each camera. Camera1 Camera7 Camera8 Camera6 Camera2 Entrance Camera Camera3 Coffee Space Camera5 Main Entrance Covered Area Course Camera4 Go to QoLT Fig. 4. Predetermined Course for Pedestrians. 10
Camera 1 Camera 2 Camera 3 Camera 4 Camera 5 Camera 6 Camera 7 Camera 8 Fig. 5. Examples of Each Camera Images. 11
Kinds of the Datasets CMU_SRD has two kinds of data. The Individual datasets are sequential images those we captured pedestrians walking with intervals. The other datasets Group are sequential images those pedestrians walking without intervals. Examples of each dataset are shown in Table 5. Table 5. Example of CMU_SRD Images ( Individual and Group ). Individual dataset Group dataset 12
5. Comparison with Other Dataset Our dataset is higher resolution than other standard dataset (PETS 2012 Benchmark Data [6] and VIPeR 1.0[7]). Table 6 shows the approximate comparison with other dataset. The resolution of pedestrians are more than twice than other standard dataset. Examples of images of each dataset are shown in Table 7. Table 6. Approximate comparison with other dataset. Dataset Name Image Size Approximate Pedestrian Size CMU_SRD 1280 x 960 [pixel] 110 x 330 [pixel] PETS 2012 768 x 576 [pixel] 35 x 100 [pixel] ViPER 1.0 48 x 128 [pixel] 48 x 128 [pixel] 13
Dataset Name Table 7. Examples of Images for Comparing Dataset. Example of Original Images Example of Pedestrian Images CMU_SRD PETS 2012 ViPER 1.0 14
6. Conclusion We will publish a new images dataset for surveillance research. This dataset can be used free for research and developments. User can download our dataset (CMU_SRD) from the following website. CMU_SRD Web Site : http://www.consortium.ri.cmu.edu/projsrd.php We hope our new high resolution dataset will be useful and have a contribution with many works related to image research. If you use this dataset in your paper/report, please read our agreement and refer this report from your paper/report. 7. Acknowledgment This data collection was supported by People Image Analysis Consortium. We also thank many participants. 8. References [1] Point Grey Official Web Site [http://www.ptgrey.com/] [2] Point Grey Official Web Site (MultiSync ) [http://www.ptgrey.com/products/multisync/] [3] Fujifilm Official Web Site [http://fujifilm.jp/] [4] HighPoint Technologies Official Web Site [http://www.highpoint- tech.com/] [5] Corsair Official Web Site [http://www.corsair.com/en- us/force- series- gt- 240gb- sata- 3-6gbps- solid- state- hard- drive] [6] PETS 2012 Benchmark Data [http://www.cvg.rdg.ac.uk/pets2012/a.html] [7] VIPeR: Viewpoint Invariant Pedestrian Recognition [http://vision.soe.ucsc.edu/?q=node/178] 15