Low-cost polarimetric imaging for surveillance

The surveillance industry has traditionally focused on the use of colour intensity images and then used computer vision methods to extract information. Deep learning methods have been demonstrated successfully but require significant computational resources. Fog and rain still present a problem to these methods. Other non-optical imaging technologies are available but the applications can be cost sensitive. Polarimetric cameras offer a solution to some of these problems. This paper presents a practical and low cost design that uses between two and four HD cameras with a wide field of view. This system has an automatic calibration stage that ensures the video frames are synchronised in time. To produce the Stoke parameters each pixel from one camera must be mapped to the others. To perform this, a homography matrix for each camera is automatically discovered and maps each video stream into the correct spatial coordinates. This attempts to use SIFT keypoint mapping but since each input image is a different polarisation state there are potentially a low number of keypoints so an additional check stage is introduced. Calibration results are presented along with example images, post process methods and feature extraction results.