CONFERENCE / ICCAIS-2026

Autonomous vehicles rely heavily on real-time perception and predictive control systems to ensure safety in dynamic environments. Existing trajectory forecasting models like Social-GAN and DESIRE offer probabilistic predictions but often struggle with maintaining consistency across highly uncertain, multi-agent scenarios, especially under occlusion or sensor noise. These limitations lead to delayed reactions or unnecessary evasive actions, affecting the system's stability and efficiency. To address these challenges, this paper presents a novel AI learning framework for proactive collision avoidance. The process begins with structured data collection using multi-modal sensors (LiDAR, RGB cameras, radar), pre-processed and fused to form a scene-aware spatial map. Then, FlowSceneGraphNet, performs spatio-temporal feature extraction by constructing a dynamic scene graph of all moving agents and decomposing their motion behaviors using relation-aware attention layers. The output of FlowSceneGraphNet is then provided as input to RandomTrajecGAN-X, a generative adversarial trajectory predictor that stochastically simulates plausible future paths for each agent based on the encoded context. The integration of these two novel algorithms enables the vehicle to anticipate collisions with higher confidence and execute evasive planning with minimal latency. Experimental evaluation using the nuScenes dataset demonstrates that our dual-algorithm pipeline reduces average prediction error by 21.4% and improves successful collision avoidance rate by 28% compared to existing baselines, proving its effectiveness in complex urban driving environments.