I'm completely new to ML and what real world applications it's suitable for. Are we at the point yet where you can train a computer to look at arbitrary images and count the number of people in it? What if it was the largely on the same background and only the number of people were changing -- for example, a camera shooting a queue of people to determine queue depth at a bus station.
A system like that would be surprisingly hard to build. The problem wouldn't be the ML algorithms - it would be just about everything else. A few things you need to solve robustly to build your counter:
1. The "same background" doesn't really exist for most cameras in most settings. Changes in illumination alone will make segmenting the background tricky. Moving objects in the scene will also be hard - think fountains and trees in the wind. Google for "foreground-background segmentation" to see some papers on this.
2. I haven't seen anyone use recent ML algorithms with less than high quality images. That may not matter, but it could matter a lot.
3. Extending recent ML algorithms to work with video at a high enough frame rate to be useful (10Hz at a minimum) may or may not be easy.
I'm sure that what you're proposing could be done. But I think that the number of small annoyances you'd hit would probably discourage most people who aren't treating the problem as a research exercise in Computer Vision.
In the scale of computer vision problems, the stationary camera case is relatively easy. It's not too hard to isolate moving objects from a background, it's not too hard to decide if an object is a person or not, and it's not too hard to keep track of an object once you've identified it. You would still have to handle overlapping people, scene illumination changes, etc, but these can be solved and have been done before.
If you would like to play with some of this stuff, take a look at OpenCV. http://opencv.org
I actually think this is quite do-able and has been for a while. Although deep learning has revolutionized object recognition, face detection has been working reasonably well for some time, e.g. your cell phone camera or Google street view face blurring.
Yes. The general task of looking at arbitrary images and labeling objects (from a set of known categories) in those images is called "detection." In fact the problem you described is easier, because there's only one category (people), and the system only needs to provide a count, rather than provide bounding box rectangles around each object (which is what the standard "detection" task entails).
Convolutional neural nets are the state of the art for this, specifically deep residual learning (http://arxiv.org/abs/1512.03385). It requires a good deal of background to understand what's going on and tune/implement the models, though, even if you just use the frameworks already out there. You probably don't even need that much data - you can probably grab pre-trained models and train them on a small additional dataset you collect.
They can definitely handle arbitrary backgrounds, although having a standard background makes the problem even easier, again.
Most deep learning computer vision algos are trained on 256x256 images, so having even larger images is just fine (you can downsample, or maybe even add up the predictions of different crops).
