Traffic routing and gateway selection for Internet flows play a crucial role in determining the performance of WMNs. Optimal routing strategies have been derived under the assumption that the traffic demands are static and known a priori, or that can be accurately predicted. However, the effectiveness of most of the existing optimization-based routing solutions is still to be demonstrated in real-world deployments. In this paper we present a practical and robust framework for load-balanced routing and gateway selection in WMNs under variable traffic loads, which relies only on real-time measurements of network statistics. Then, we design, implement and evaluate two simple adaptive strategies to automatically select network paths and gateways for flows that dynamically arrive and depart, with the goal of achieving a high network utilization and improved load balancing. We have developed a full-fledged implementation of the proposed mechanisms as an extension of the OLSR protocol, and conducted an experimental study in a trial outdoor mesh network using video streaming traffic. Collected measurements show that our scheme can significantly increase the number of video sessions that can be supported with acceptable quality compared to traditional shortest path routing.