| Summary: | 碩士 === 國立成功大學 === 測量及空間資訊學系 === 106 === Being an indoor positioning technique used in GNSS-denied environment, Pedestrian Dead Reckoning (PDR) is well-known for its building independent positioning algorithm. It effectively refrains from the influence due to surrounding environment, and locates the device using only self-contained sensors such as tri-axis accelerometer, tri-axis gyroscope and tri-axis magnetometer. By applying behavior recognition, step detection, step length estimation and heading estimation, the user’s walking dynamic will be modeled; then, the user’s location and the travelled trajectory can be calculated. In most researches related to PDR, the attitude of portable devices embedded with various sensors are assumed to be fixed with respect to the user. However, this assumption will be reasonable only when the device is placed in pocket, fixed on foot, fixed on belt, etc. For above situation, the misalignment between sensor frame and pedestrian frame should be constant or merely regularly varying. However, when the handheld mode PDR is applied, there will be various unpredictable behaviors and vibrations cause irregular misalignment. Thus, how to effectively and precisely estimate the misalignment in high hand dynamic situation forms the target of this research. Thanks to the development of technology, computer vision and Convolutional Neural Network (CNN) have become popular in recent years. These technologies have been successfully and widely implemented in many applications such as image recognition, obstacle detection and motion tracking. To enable the usage of sensors in non-constrained ways, computer vision-based and CNN-based methods are used for misalignment correction in this research. In the result, the analysis of misalignment estimation will be provided. Besides, the corrected PDR trajectory is also given for validating the effectiveness of misalignment correction.
|
|---|
