Computer Vision-assisted Battery-free RFID Systems for Object Recognition, Localization and Orientation

Publication Type:
Issue Date:
Full metadata record
Battery-free radio frequency identification (RFID) is a promising technique in Internet of Things (IoT) applications that use wireless signals to identify a physical object from its attached RFID tag. Compared to the existing barcode identification systems, RFID can still work in the non-line-of-sight (NLOS) scenarios that some obstructions block the identifier. Recently, many researchers start regarding each RFID tag as a battery-free sensor, whose indicator is the backscatter signal fingerprint reported by an RFID reader. Since the sensor could sense the change in the position and orientation of an RFID tag relative to a reader antenna as well as surroundings, a variety of battery-free RFID sensing systems are proposed for object localization, direction tracking, material recognition, human breathing/heartbeat rate assessment, liquid leakage detection, etc. However, some technical challenges still remain to be addressed in these purely RFID-based systems. This thesis introduces computer vision (CV) techniques into RFID systems to minimize the impact of RF phase periodicity and multipath interference. In the thesis, three categories of CV-assisted battery-free RFID systems for object recognition, localization and orientation are designed, and the main contributions include: 1) This thesis presents RF-Focus, a CV-assisted system that recognizes moving RFID-tagged objects within the region of interest and tracks their trajectories in multipath environments. To achieve RF-Focus, novel RSSI/RF phase-distance models with additional multipath terms compared to traditional models are proposed to characterize the impact of multipath interference, and thereby a dual-reader-antenna solution is designed to deal with it. Moreover, the multipath terms in RSSI and RF phase can be leveraged to clean the phase shift caused by frequency-dependent RFID hardware characteristics in RF phase. After that, an innovative fusion algorithm is designed to match position proposals outputted by a 2D camera and the cleaned RF phase for object recognition. In the experiments, RF-Focus achieves 91.67% ROI object recognition in multipath environments when simultaneously tracking five moving objects. 2) This thesis proposes RF-MVO, a CV-assisted system that locates stationary RFID tags in 3D space without driving a platform carrying reader antennas along a predefined trajectory or pre-deployed track. To achieve RF-MVO, a 2D camera is affixed to reader antennas. A fusion model is designed to fuse camera trajectory in the camera view with depth-enabled RF phase to achieve real-world trajectory transformation and tag DOA estimation. On this basis, a novel 3D localization is proposed, which could avoid consuming huge computations to search for all possible regions. In addition, a joint optimization algorithm is designed to accelerate RFMVO and improve its estimation accuracy. Finally, this thesis introduces horizontal dilution of precision widely used in satellite positioning systems to find out the optimal localization result. The experiments show that RF-MVO achieves 6.23cm localization accuracy in 3D space. 3) This thesis proposes RF-Orien3D, a CV-assisted system that leverages the variation of each tag radiation pattern in a two-RFID-tag array to estimate a labelled object’s spatial directions (i.e., azimuth and elevation) in multipath environments. To achieve RF-Orien3D, this work proposes novel RSSI/RF phase-distance models when tag mutual coupling and multipath interference both occur. In the models, one variable to be estimated is tag radiation pattern, which is simulated by building a two-tag array from a 2D image; another is modulation factor, which is estimated using RFID fingerprints in non-coupling and coupling in free space. On this basis, a convolutional neural network (CNN)-based method is proposed by simulating all multipath impacts on RFID fingerprints based on the proposed fingerprint models to pre-train a CNN and then collecting measured data to fine-tune the CNN for 3D orientation. In the experiments, RF-Orien3D achieves median angle errors of 29° and 11° in azimuth and elevation.
Please use this identifier to cite or link to this item: