Human Presence Detection
The following three videos are based on a compilation of images taken by the USB web cam every 2 minutes throughout a day. The first video is the raw images compressed into an AVI file. The second is a simple difference algorithm where the previous frame is subtracted from the latest frame. The final video is based on a simple algorithm
Xt = (1-alpha)*Xt-1 + alpha*Xt
Frames were computed for alpha values of 0.5 and 10 and the resulting video is the difference of these two X_t computations.
The second set of videos are from captures that used a combination of periodic and motion based captures.
Another format of video has been added for both the periodic and motion captures. This new video displays the original images, frame to frame difference, and X_t with substitution next to each other for easy comparison. The X_t substitution works by converting each X_t frame to a gray scale image. Then thresholding is applied pixel by pixel to identify which pixels have experienced a significant enough change. These pixels are then replaced with the pixel of the same position from the original image. Through this process it became clear that the original algorithm was mostly responding to new dark items in the room.
The following plots show what percentage of the pixels exceed each threshold. Using this information the threshold of 1 was chosen as there did not seem to be a significant change with higher thresholds.
--Yoavfreund 17:52, 13 May 2010 (UTC) Threshold 0 seems to be the most informative. Can you change the horizontal axis so it represents the time of day, rather than the frame number?
The camera being used comes equipped with a motion-triggered image capture system. By combining this system with periodic captures, we hope to find a higher number of captures with people within the frame. A second day of photos was collected using this combination. The videos for applying the same basic algorithms to this new data set can be found on the same page.