@丕子‘s comment reminded me to broaden the review area to cover multi-objects detection. Apologise for not making clear what is about to review in the second part, as a continue review of people counting using openCV, as in part 1 the review did not tell a clear content structure.
@zhijie thanks for your kind wishes.
Challenges for people counting using openCV
In part 1 the review mentioned some difficulties to apply openCV for people counting, including camera positions, specific detection feature (faces) and overlapping objects.
There are some other challenges as well – multiple people detection and tracking. This even concerns previous challenges, like the object overlapping. Someone suggested to use infrared photodetectors to avoid such issue, while that does not make sense here in using openCV and camera.
Methods of people counting using openCV
Andrey‘s suggestions gave three ideas from simple to advance. Simply enough, it was advised to calculate differences between successive frames of video stream and this detected line-crossing event by determining motion masks. A little more complicated, the idea was to use ‘running average‘ approach, and the most advanced one, was to integrate background subtraction technique to enhance the accuracy. The running average can be reached here openCV cvRunningAVG. However, there is still a problem – what if people gather together as a group then enter the room? In other words, how to keep tracking objects that may be combined and separated, unpredictably.
Counting grouped people using openCV
A short paper was recommended by someone in overstackflow.com, about human tracking by fast mean shift model seeking. Due to the limitation of time, the post is not demonstrating the details, but a picture is presented for demonstration. The paper can be reached via the link provided on the bottom of the post. This paper shows tracks behind people’s movements, grouping and separating have little influence on it.
Other ways of people counting using openCV
Other approaches were also introduced to deal with people group gathering problem. Someone (sorry I forgot the source) used track ID to mark different states of people. Every individual was given a unique track ID, which was binded with pixel-based contours. When two contours emerged together, the two track IDs would follow the common contour, until these people entered the room and were counted by the system.
A short summary
Neither people counting using openCV in part 1 and 2 is posted with some new cutting edge techniques, with which the people counting could be overwhelmingly improved. Rather, the state of people counting using openCV is very briefly introduced, with some concrete methods for people counting use. Wish these could be sparking some new thoughts.
Looking back the methods for people counting using openCV, some achievements have been made, especially in terms of counting a general number of people within specific settings. However, the future developments still face some problems, such as accurate grouped people counting, dwelling people counting and, probably, super density people counting – will it be possible to do that by counting heads?
Any comments are warmly welcome, including criticisms, as this review is widely open to anyone who is interested in people counting using openCV.
5 Comments On “ A review of people counting using openCV part 2”