This paper proposes a vision-based control scheme for safe robot navigation in crowded environments. Unlike traditional methods relying on LiDAR or laser rangefinders, our approach leverages an RGB-D camera to capture rich visual information about the surroundings, allowing for a more comprehensive understanding of the scene. We address the challenge of predicting human motion in dynamic environments by combining a vision-based human detection module with a crowd prediction module. This allows the robot to anticipate potential collisions and generate safe motions. Additionally, we introduce an adaptive camera control strategy to enhance human detection performance by following their movement within the field of view of the camera. The proposed control scheme utilizes Control Barrier Functions (CBFs) to enforce safety constraints. By incorporating information about both robot-human relative position and velocity, CBFs ensure collision avoidance even in dynamic scenarios. The effectiveness of the method is evaluated by comparing the performance of different human detection algorithms, and by demonstrating the benefits of the adaptive camera control strategy and the overall safety achieved through the proposed vision-based control scheme.