Object recognition in image is still in its primitive stage in academic research. Yet, there are already efficient algorithms available for matching image patterns, such as SIFT. Already, companies are started to make real image search usable, such as TinEye.
Present relevant ads while video is being played online.
As video is played, match the videos frames against a huge library of product image patterns. Think about Coke, Nike, iPhone. Whenever such a product appears in the video frame, an image match can be found, which then triggers a text-search by the product key work, in the ads network, like the AdWords network. The found ads text is then displayed, in a very non-intrusive fashion, on the right-hand-side panel, for the viewer to click, just in case, the viewer is not 100% focused on the contents of the video, which is highly likely.
Quick note that, the image search does not need to happen in real time. Vast preprocessing can be done overnight for all videos and tag the ads into the video player.