Please enable JavaScript in your browser.

fltech - 富士通研究所の技術ブログ

富士通研究所の研究員がさまざまなテーマで語る技術ブログ

Introducing Camera Angle Change Detection

Hello, I'm Yoshie Kobayashi from our Artificial Intelligence Laboratory. Today, I would like to introduce an exciting new AI core engine - Camera Angle Change Detection – which is now available from Fujitsu Kozuchi .

This blog is the fourth part of a five-week blog series. The last post, " Introducing Multi-Camera Tracking " is here !

Camera Angle Change Detection is a technology that can immediately detect slight camera angle changes. It is one of Fujitsu Kozuchi’s AI core engines, which enables the fast testing of cutting-edge AI technologies developed by Fujitsu.

Security cameras and other fixed cameras are set up in many public places like shops, inside factories, stations, and airports. Recently, there's been more use of image recognition AI to analyze the footage from these fixed cameras. However, if the direction of the fixed cameras is shifted due to a shock or unexpected movement, or if there is even a slight change in the shooting range, the accuracy of the image recognition AI decreases. If the accuracy decreases, it is necessary to correct the definition information of the shooting location or retrain the AI. But if you can't detect the change, you can't correct it. To overcome this issue, we have developed Camera Angle Change Detection.

Benefits of Camera Angle Change Detection and how to use it

The principal value of Camera Angle Change Detection is that it can immediately detect slight camera angle changes, preventing operation with an altered range and enabling the stable operation of image AI.

Camera Angle Change Detection can not only detect changes in the camera angle, but also detect changes in the range. It can detect camera rotation, enlargement/reduction of the range, and changes in layout such as machines in a factory or product shelves in stores. Detection is carried out using only two images from the camera footage. The time it takes for detection is about 500msec (detection time only). The following is an example from inside a factory. It detected slight changes between the two images and detected camera rotation.

In addition, Camera Angle Change Detection is easy to apply. Since it does not require a GPU (a processor specialized for image processing) to make a judgement, it can be applied to any environment. As Python is used for the programming, it can be easily combined with other engines as many of them are also programmed using Python.

Features of Camera Angle Change Detection technology

How does it work? Camera Angle Change Detection detects camera angle changes by calculating the rate of feature changes from the images, and estimates these changes using a change matrix. By finding the same features from the images before and after the change, and estimating the amount of change through optical flow, we can determine the change matrix. We calculate the change rate from the change matrix, and identify whether the change that occurred according to that change rate is rotation, enlargement/reduction, or layout change. Therefore, it requires very little effort to apply, as it does not require pre-training like DeepLearning or a GPU.

Interested in testing the Fujitsu Kozuchi?

Our technology detects changes in appearance caused by important factors in the operation of AI, such as camera rotation and enlargement/reduction. This can prevent AI from making false detections. Since there is no need to check each camera manually, it can also reduce the amount of work. It is a technology that greatly contributes to the stable operation of AI.

For a demonstration or to test our Camera Angle Change Detection, please contact us here :