Assistive rehabilitation gloves are critical for patients with hand dysfunction, but existing systems suffer from poor portability, and reliance on subjective therapist assessments. This letter presents a compact, embedded rehabilitation glove system based on the STM32F429 microcontroller, enabling personalized and precise rehabilitation training. A pruned cross-entropy support vector machine (SVM) algorithm is proposed to realize real-time movement recognition through visual data (320×240 pixels, 1 fps), overcoming the deployment limitation of complex networks on embedded hardware. A rehabilitation index α is defined with cross-entropy to quantify movement similarity with pre-stored standard movements is utilized, driving adaptive variable impedance control(VIC) to adjust movement stiffness. Experimental results show the system achieves 83.8%–94.3% recognition precision, a maximum grasping force of 1.4 N (Newton), and effective completion of dexterous tasks (e.g., egg grasping, cylinder manipulation). This portable solution realizes data-driven rehabilitation and addresses the gap between professional therapy and home-based training.