Research Of Multi-Sensor Fusion Based Outdoor Object Location Recognition On Smartphones

It has been a research hotspot in the field of mobile Internet to leverage various sensor data from smartphones for providing location-based services. However, most of the existing outdoor localization methods can only acquire users' physical coordinates, but cannot recognize the logical meaning of the object location that users are interested in. In addition, the existing outdoor object location recognition methods usually use content-based image retrieval technologies or only rely on non-image information such as orientation and angle, which are hard to balance the recognition accuracy and time complexity.In the face of these challenges, a novel outdoor object location recognition method is proposed in this paper. Object location can be achieved by taking a picture of the object and fusing other sensor information to build the location fingerprint and matching it with a pre-built location fingerprint database. To meet the goal of user friendliness,high accuracy and quick response, firstly, we present an object localization method based on the perspective projection model. Only a simple user action is required to not only assist in object localization, but also normalize the vertical viewing angle and thus reduce the randomness of fingerprints. Secondly, we propose an "object-centric" fingerprint search scheme, which can make full use of smartphone sensor data, and then determine a more accurate search space. Finally, a fingerprint matching method is proposed to combine visual fingerprint and non-visual fingerprint for improving the recognition accuracy and reducing the response time. Based on these methods, we implement an outdoor object location recognition system, and collect location fingerprints in the BUPT campus to construct an experimental dataset for evaluating the availability and the performance of the system. It is verified that our method achieves satisfactory results.