Eye-gaze-based control has emerged as a critical assistive technology for individuals with severe motor impairments who are unable to operate conventional wheelchair interfaces. Existing systems often rely on single-channel computer vision approaches that suffer from user variability, lighting changes, and instability, or depend on expensive infrared eye-tracking hardware. This paper presents a hybrid eye-gaze-controlled wheelchair navigation framework that integrates geometric Eye Aspect Ratio (EAR)-based blink detection with convolutional neural network (CNN)-based gaze direction classification through a temporal stability filtering mechanism. Using facial landmarks, a user-calibrated CNN trained on a custom dataset analyzes normalized eye regions extracted from real-time webcam frames. Concurrently, EAR values derived from facial landmarks robustly detect intentional long blinks, which act as high-priority stop commands. A multi-frame decision buffer enforces temporal consistency, allowing only stable and high-confidence gaze interpretations to be translated into navigation commands. Experimental results show that the proposed system achieves an eye-movement direction recognition accuracy of 96.2%, along with precise blink detection and smooth command execution using only a standard RGB webcam and commodity hardware. These results demonstrate that the proposed approach effectively combines personalized training, temporal stability, and CNN-EAR fusion into a single low-cost and reliable assistive navigation system.