Human-Robot Collaboration in industrial context requires a smooth, natural and efficient coordination between robot and human operators. The approach we propose to achieve this goal is to use on-line recognition of technical gestures. In this paper, we present together, and analyze, parameterize and evaluate much more thoroughly, three findings previously unveiled separately by us in several conference presentations: 1/ we show on a real prototype that multi-users continuous real-time recognition of technical gestures on an assembly-line is feasible (≈ 90% recall and precision in our case-study), using only non-intrusive sensors (depth-camera with a top-view, plus inertial sensors placed on tools); 2/ we formulate an end-to-end methodology for designing and developing such a system ; 3/ we propose a method for adapting to new users our gesture recognition. Furthermore we present here two new findings: 1/ by comparing recognition performances using several sets of features, we highlight the importance of choosing features that focus on the effective part of gestures, i.e. usually hands movements; 2/ we obtain new results suggesting that enriching a multi-users training set can lead to higher precision than using a separate training dataset for each operator.