Our research tackles a complex-networks mining task known as community detection, in a strongly decentralised and distributed context: we consider the case of opportunistic networks comprised of small wireless connected devices, communicating peer-to-peer. We propose to combine the Think-Like-a-Vertex graph processing paradigm with leader and seed-based community detection methods, suitable for decentralisation. We propose a global working principle which we implemented through three algorithms dealing with three different aspects of the community detection task: detection of disjoint communities on static graphs, detection of overlapping communities on static graphs and the case of dynamic graphs. We present these algorithms, together with an experimental study on benchmarks and real data to assess the quality of the results and to compare them with state-of-the-art methods. We also consider, in the specific case of an opportunistic mobile network comprised of smart communicating clothing, a task of pathfinding towards target, an unknown member of the social network. We propose a recommandation strategy exploiting the graph's community structure, designed and evaluated through an algorithm.