This paper proposes a decentralized bilevel optimization framework for real-time vehicle-to-vehicle (V2V) collision avoidance. To balance collision risk, velocity consensus, and speed efficiency, the upper level uses Particle Swarm Optimization (PSO) to adaptively adjust risk-weight parameters. The lower level solves a constrained Quadratic Programming (QP) problem to compute safe and efficient velocities while satisfying physical and safety constraints. Each vehicle autonomously identifies its neighbors using local sensing and adjusts its trajectory based on a composite risk function. This two-tier structure enables context-aware real-time coordination without the need for centralized control. The simulation results demonstrate that the proposed method achieves smooth velocity alignment and significantly reduces the risk of collisions over time between multiple autonomous vehicles. The decentralized architecture provides scalability and resilience, making it suitable for dense and dynamic traffic environments. In general, this work presents a practical, flexible, and computationally efficient solution for collaborative vehicle motion planning under uncertainty.