Managing collision avoidance between multiple drones in confined spaces represents a critical challenge in autonomous systems. While existing approaches have made progress, they often struggle to balance immediate safety requirements with long-term performance optimization in controlled environments. Traditional collision avoidance methods provide reliable safety guarantees but lack adaptability, while Deep Reinforcement Learning (DRL) approaches offer learning capabilities but suffer from stability issues in dense scenarios. To address these limitations, we propose a novel hybrid framework that integrates geometric Time-to-Collision (TTC) calculations with Deep Deterministic Policy Gradient (DDPG), leveraging their complementary strengths. Our approach bridges the gap between reactive safety mechanisms and learning-based optimization, providing both robust collision avoidance and optimal trajectory planning. Experimental results demonstrate significant improvements over existing methods, achieving collision rate reductions to 7% and 19% for 5-drone and 10-drone scenarios respectively, while accelerating learning convergence by 31% and enhancing stability with 53% lower reward variance. The framework exhibits particularly robust performance in high-density environments, maintaining consistent performance despite increasing system complexity. These results establish a new benchmark for safe and efficient multi-drone control systems in real-world applications.