To mining video data, first we need to understand how it is different from other types of data. Video is essentially a sequence of image frames. Video data of a normal camera has about 30 image frames per second. Each image frame is identical to a still image.
The first step of video mining is segmentation. This is similar to the first step in image mining. The goal of segmentation is separating objects (such as vehicles) from their background. In the following pictures, the left is the full picture, and the right is the background.
The second step of video mining is object tracking. For each object, a bounding box is calculated and the centroid of this box is located. Then the position of centroid is tracked from frame to frame. The change of such centroid gives us the movement of the object.
The ability of tracking object allows us to estimate traffic flow through a interaction between, say 9 am to 10 am.
For more information on how to mine traffic data, please see this paper:
Chen, Shu-Ching, et al. "Multimedia data mining for traffic video sequences." Proc. of International Workshop on Multimedia Data Mining (MDM/KDD’2001). 2001.