Shared control improves robotic obstacle avoidance by combining human guidance and autonomous intervention, but existing methods struggle to balance them adaptively. To address this, we propose a shared control framework that dynamically adjusts control authority based on real-time obstacle proximity. The framework combines Weighted Quadratic Programming (WQP) with a sigmoid-based adaptive weighting mechanism for smooth transitions between operator-guided and autonomous collision-free trajectories. Autonomous trajectories are generated by a neural network policy trained on collision-free demonstrations from an enhanced A* algorithm. A comprehensive metric, Obstacle Avoidance Efficiency (OAE), is proposed to evaluate trajectory safety, efficiency, and risk exposure. Experiments validate the framework in two scenarios: static obstacle and suddenly placed obstacle. The experimental results demonstrate that our method outperforms direct operation, manual switching, and fixed-weight methods. It achieves higher success rates and better obstacle avoidance performance, effectively balancing human intent and safety.