The real limitation of using unobtrusive observation to collect data is in the perceptual distortions made by the observer. We tend to see what we expect to see. Just being aware of the potential theoretical frameworks that others may have developed with which to explain the phenomenon under investigation is enough to prime our perceptual mechanisms and bias the data collected.
All manner of techniques been developed to try and limit the potential for perceptual distortion - interrater reliability, literature searches for disconfirming data, peer and mentor checking. These are all worthwhile efforts to improved the objectivity of the observations, but the limitations remain a concern.
In addition to Peter's comment above, which is totally accurate, please have a look at the ICC/ESOMAR INTERNATIONAL CODE ON MARKET AND SOCIAL RESEARCH (http://www.iccindiaonline.org/policy_state/esomar.pdf), specifically at Article 6:
Article 6 – Recording and observation techniques Respondents shall be informed before observation techniques or recording equipment are used for research purposes, except where these are openly used in a public place and no personal data are collected. If respondents so wish, the record or relevant section of it shall be destroyed or deleted. In the absence of explicit consent respondents' personal identity shall be protected.
If you are sure that your collection of data does not contravene the above, then you're okay.
The most effective approaches to collecting unobtrusive data use systematic protocols that allow you count things. Note that this is quite different from things like ethnography or participant observation, where your presence in the setting is an "obtrusive" aspect of the research.