Augmented Reality (AR) has been used to facilitate surgical guidance during External Ventricular Drain (EVD) surgery, reducing the risks of misplacement in manual operations. During this procedure, the pivotal challenge is the accurate estimation of spatial relationship between pre-operative images and actual patient anatomy in AR environment. In this research, we propose a novel framework utilizing Time of Flight (ToF) depth sensors integrated in commercially available AR Head Mounted Devices (HMD) for precise EVD surgical guidance. As previous studies have proven depth errors for ToF sensors, we first conducted a comprehensive assessment for the properties of this error on AR-HMDs. Subsequently, a depth error model and patient-specific model parameter identification method, is introduced for accurate surface information. After that, a tracking procedure combining retro-reflective markers and point clouds is proposed for accurate head tracking, where head surface is reconstructed using ToF sensor data for spatial registration, avoiding fixing tracking targets rigidly on the patient's cranium. Firstly, $7.580\pm 1.488 mm$ ToF sensor depth value error was revealed on human skin, indicating the significance of depth correction. Our results showed that the ToF sensor depth error was reduced by over $85\%$ using proposed depth correction method on head phantoms in different materials. Meanwhile, the head surface reconstructed with corrected depth data achieved sub-millimeter accuracy. Experiment on a sheep head revealed $0.79 mm$ reconstruction error. Furthermore, a user study was conducted for the performance of proposed framework in simulated EVD surgery, where 5 surgeons performed 9 k-wire injections on a head phantom with virtual guidance. Results of this study revealed $2.09 \pm 0.16 mm$ translational accuracy and $2.97\pm 0.91 ^\circ$ orientational accuracy.