JAIST Multi-View Surveillance (MVS) Video Database

 
From October 2010, a new project is launched to study the autonomous report generation for multiple view surveillance environment. This research targets at efficient management and retrieval of media data in future large-scale surveillance system. Especially, we aim at developing a system to automatically generate an online/offline field reports to satisfy various user preferences, based on contextual informations.
   *This research is supported by the 2010 Research Grant-in-Aid (Start-up Support).

In this page, we present a database, namely the JAIST-MVS database, which collects the video data that we need to study people tracking, abnormality detection and report generation.

Public Distribution

With the permission of the life-science commitee, we are able to make the data available to the researchers in computer vision community, after signing a user agreement. We will only provide the low resolution version of the database, which will be available through network downloading. The high quality version and all other necessary information may be accessed by sending an Email to Dr Fan CHEN at chen-fan AT jaist.ac.jp , giving the names of the researchers who wish to use the data and their main purposes.

Acquisition Setup

This video database was acquired by eight high definition surveillance cameras (Sanyo VCC-HD2300 with wide-angle lens Fujinon YV2.8x2.8SA-2) that were setup in our lab room.
vcc-hd2300 YV2.8x2.8SA-2
Sanyo VCC-HD2300 [Datasheet] Fujinon YV2.8x2.8SA-2[Datasheet]


All videos were taken in a rectangle lab room, where the action zone is around 3.5m x 4.5 m.

action-zone



Here is an image to provide an initial idea of each camera view and its effictive area.

calib results

Action Lists

With the permission of the life-science commitee, we organized one video acquisition activity. All videos were recorded at around 30FPS. Since the cameras accept no external trigger signals, all videos were only software synchronized. The users may need to adjust the time offset of each video. In this acquisition, we include three groups of video data into the JAIST-MVS database:

 1. Single actions

In this subset, we ask each person to perform an action in the middle of the room. We collect eight types of single-person actions from five different persons.
  A0S: Normal walking
Enter the room, walk for two rounds and exit the room.

thumbnail a0s

  A1S: Drunkard walking
Enter the room, drunkard walk for two rounds and exit the room.

thumbnail a1s

  A2S: Sneak walking
Enter the room, sneak walk for two rounds and exit the room.

thumbnail a2s

  A3S: Object taking/returning
Enter the room, take a bag away from the central table, and then exit the room.
Enter the room again and put the bag back.

thumbnail a3s

  A4S: Object smashing with a hammer
Act as crushing an unreal object in the middle of the room, with a hammer.

thumbnail a4s

  A5S: Peek into a car
Assume that there is a car and peek into the window。

thumbnail a5s

  A6S: Falling down
Act as a patient, who slowly falls down due to heart attack.

thumbnail a6s

  A7S: Lock picking
Assume that there is a door in the middle of the room. Stop by in the middle of the room, place a knee on the mattress, and take out an electrical driver.

thumbnail a7s

 2. Group actions

In this subset, we ask several persons to perform an action in the middle of the room. We collect eight types of actions.
  A0G: Group walking
More persons are walking freely, with leaving and entering the room

thumbnail a0g

  A1G: Group walking with interaction
More people are walking freely. Time by time, two persons walk closer, have a short chat, and then separate.

thumbnail a1g

  A2G: Object dropping
The first one drops something, and the second one picks it up and leaves. Other people keep walking.

thumbnail a2g

  A3G: Bag exchange
Two persons meet each other and exchange their bags/suitcases. Other people keep walking.

thumbnail a3g

  A4G: Bag stealing
The first one leaves his suitcase for a while, and the second one takes it away. Other people keep walking.

thumbnail a4g

  A5G: Fighting
Two people start to fight and all other people stop by and circulate these two persons.

thumbnail a5g

  A6G: Falling down due to external forces
A push B, which makes B fall down. A leaves the place quickly. Other people start to check B's situation. B stands up. All persons continue normal walking.

thumbnail a6g

  A7G: Falling down due to internal reason
Person A falls down, and Person B comes closer to Person A for checking the situation and then start to call other people. Other people start to circulate them.

thumbnail a7g

 3. Long videos of free combination of actions.

In this subset, we ask several persons to perform a free combination of actions. According to the density of abnormal events, we collected three videos.
  Overall 0: High abnormal event density

thumbnail overall dense

        Ground-truth data of people tracking from manual labelling. [Updated Jul. 23, 2011]

Demo video

  Overall 1: Medium abnormal event density
(Thumbnail Omitted)
  Overall 2: Sparse abnormal event density
(Thumbnail Omitted)

 4. Some People Tracking Results. [Updated Jun. 14, 2011]

We show here some people tracking results based on the above database.
  Results on the video of Overall 0: High abnormal event density

Demo video

  Results on the video of Overall 1: Medium abnormal event density

Demo video

  Results on the video of Overall 2: Sparse abnormal event density

Demo video

   
Note: All demo videos are encoded in xvid mpeg 4 format. If you could not replay the demo video, please download the latest vlc media player for your os or other proper players. Get VLC Media Player
   







Maintained by chen-fan AT jaist.ac.jp. Created on 2011-4-7. Last update 2011-7-23
Valid HTML 4.01 Transitional