Abstract: Sky polarization field positioning technology, characterized by full autonomy, anti-interference capability, and no cumulative error, offers a novel solution for positioning in situations without satellite signals. To address the issues in current imaging-based polarization positioning devices, which often use fisheye lenses requiring complex full-field calibration and suffer from polarization effects, a positioning device based on specular reflection is designed and implemented with system errors compensated using a neural network approach. The device consists of a polarization camera, a narrow-field lens, multiple mirrors, and a north-seeking instrument. During measurement, scattered light from the sky first strikes the mirrors, generating reflected light, which then passes through the narrow-field lens before reaching the image sensor of the polarization camera. A model is established to derive the polarization state of the incident light from the polarization state of the reflected light. Based on this model, the device calculates the solar vector from the polarization state of the incident light at each observation point and, in combination with astronomical almanac and time information, achieves geolocation functionality. Additionally, a neural network model is employed to compensate for system errors. Outdoor static experiments show that the root mean square errors for longitude and latitude positioning are 0.2596° and 0.3362°, respectively, with a maximum error not exceeding 0.6°. The results demonstrate that the system can provide real-time and stable geographic positioning information in the absence of satellite signals.