Joint limits avoidance is a crucial issue in sensor- based control. In this paper we propose an avoidance strategy based on a low-level data fusion. The joint positions of a robot arm are considered as features that are continuously added to the control scheme when they approach the joint limits, and removed when the position is safe. We expose an optimal tuning of the avoidance scheme, ensuring the main task is disturbed as little as possible. We propose additional strategies to solve the particular cases of unsafe desired position and local minima. The control scheme is applied to the avoidance of joint limits while performing visual servoing. Both simulation and experimental results illustrate the validity of our approach.