Driver drowsiness is a major reason for road accidents, especially on unstructured or under-marked roads, where conventional Advanced Driver-Assistance Systems (ADAS) generally fail to work. The paper proposes a vision-based real-time autonomous parking system triggered by driver drowsiness detection. The system integrates SegFormer-B0 for efficient road segmentation, Sobel edge detection for boundary amplification, and YOLOv8x-seg for real-time object segmentation and recognition. When drowsiness is identified, the pipeline computes the distances to neighboring obstacles and road boundaries in pixel-space Euclidean units. When a safe parking spot is identified in accordance with deterministic logic rules, the car carries out an autonomous parking procedure along a smooth cubic Bézier curve path, with the addition of a simple kinematic motion model to make maneuvering realistic and stable. The Bézier curve dynamically adapts its control points to respond to real-time movement of obstacles. Extensive testing on structured and unstructured road datasets demonstrates the system provides stable perception, safe decision-making, and uninterrupted autonomous parking execution with mid-range GPU hardware real-time performance (~15–20 FPS). This integrated approach proves the feasibility of combining advanced perception, kinematic modeling, and trajectory planning for enhanced autonomous emergency parking in real-world driving conditions.