The amount of data generated by surveillance cameras is vast due to the increasing number of installed cameras in city environments. Since the human capacity of analyzing such a huge volume of data is limited, machine learning techniques capable of assisting human operators are desirable for many scenarios. A common task that usually overwhelms human operators is analyzing features and behaviors of people groups in everyday situations. This work proposes a modular framework flexible enough to extract features and behaviors to analyze low-density clusters useful in real-world applications. A proof-of-concept is implemented to show the viability of the proposed modular framework in composing approaches to analyze clusters. Our experiments show that the implementation behaves as expected and can process videos with different resolutions, achieving reasonable frame rates.