Читать книгу Position, Navigation, and Timing Technologies in the 21st Century - Группа авторов - Страница 55

Fingerprinting techniques refer to algorithms that estimate the location of a person or object at any time by matching real‐time signal measurements with unique location‐specific “signatures” of signals (e.g. Wi‐Fi RSSI). Typically, fingerprinting can be performed analytically or empirically.

Analytical fingerprinting, for example, RSSI‐based, involves using propagation models such as the radial symmetric free‐space path loss model to derive the distance between a radiating source and a receiver by exploiting the attenuation of RSSI with distance. Unfortunately, this simplistic model is rarely applicable in indoor environments, where the signals do not attenuate predictably with the distance due to shadowing, reflection, refraction, and absorption by the indoor building structures. Therefore, other models have been proposed, such as the Indoor Path Loss Model [58] and the Dominant Path Model [59], which takes into account only the strongest path, which is not necessarily identical to the direct path.

Empirical fingerprinting is more commonly used in various indoor localization techniques due to the difficulty in analytically modeling unpredictable multipath effects. There are typically two stages involved in such empirical location fingerprinting: an offline (calibration) stage and an online (run‐time) stage. The offline stage involves a site survey in an indoor environment, to collect the location coordinates/landmarks/labels and strengths (or other features) of signals of interest at each location. This procedure of site survey is time consuming and labor intensive. However, such a survey can account for static multipath effects much more easily than with analytical fingerprinting (although dynamic effects, e.g. due to different number of moving people are still problematic and can cause variations in readings for the same location). Several public Wi‐Fi APs (and also cellular network ID) databases are readily available [60–63] that can somewhat reduce survey overheads for empirical‐fingerprinting‐based indoor localization solutions; however, the limited quantity and granularity of fingerprint data for building interiors remains a challenge. In the run‐time stage, the localization technique uses the currently observed signal features and previously collected information to figure out an estimated location, with the underlying premise that the locations of interest each have unique signal features.

RSSI‐based empirical fingerprinting is used extensively in several indoor localization techniques. Many RSSI‐based fingerprinting solutions aim to utilize the existing infrastructure to minimize costs, for example, Wi‐Fi APs [65] and GSM/3G/4G cellular networks [66], while a few approaches advocate for custom beacon deployment for RF signal generation to support RSSI‐based localization [67, 68]. A GSM cellular network RSSI‐based indoor localization system is presented in [66]. Indoor localization based on a cellular network is possible if the locale is covered by several base stations or one base station with strong RSSI received by indoor cellular devices. The approach uses wide signal strength fingerprints, which includes the six strongest GSM cells and readings of up to 29 additional GSM channels, most of which are strong enough to be detected but too weak to be used for efficient communication. The additional channels help improve localization accuracy, with results showing the ability to differentiate between floors in three multi‐floor buildings, and achieving median within‐floor accuracy as low as 2.5 m in some cases. Typically, Wi‐Fi is the most common signal type used in RSSI‐based fingerprinting approaches. Figure 37.5 illustrates how different locations often are covered by different Wi‐Fi APs or have different signal strength characteristics for the same Wi‐Fi AP, as can be seen from the two plots in the figure, allowing for unique fingerprinting and consequently, localization [64].

RADAR [65] was one of the first approaches to use Wi‐Fi RSSI for indoor localization. An offline phase was used to measure Wi‐Fi AP signal strengths across locations. Signal attenuations due to floors, walls, and other obstructions were also modeled, to improve the accuracy of the system to about 2–3 m in some cases. Several other indoor localization approaches use a similar strategy, for example, PlaceLab [69] and Horus [70]. Probabilistic (Bayesian‐network based) methods were employed in [71, 72] to improve the correlation between locations and Wi‐Fi RSSI during fingerprinting. In [73], a neural network classifier was used for Wi‐Fi RSSI‐based location estimation, with a reported error of 1 m with 72% probability. Wi‐Fi RSSI fingerprinting has also been widely used in the area of mobile robotics, to determine the location of a mobile robot assuming the availability of inputs from robot‐mounted sensors. A Bayesian robot localization algorithm was proposed in [74] that first computed the probability of a robot’s location based on the RSS from nine Wi‐Fi APs, and then as a second step exploited the limited maximum speed of mobile robots to refine the results (of the first step) and reject solutions with significant change in the location of the mobile robot. Depending on whether the second step is used or not, 83% and 77% of the time, mobile robots could be located within 1.5 m in their studies. A challenge in these fingerprinting‐based techniques is that two distant locations in an open indoor space may have similar signal fingerprints, which can be very problematic during localization [75]. Fingerprinting using the RSS of FM radio signals was proposed in [76]. As FM signals do not carry any timing information, which is a critical factor in range calculation using ToA, TDoA, and AoA methods, RSS‐based fingerprinting is the most viable approach when using FM signals for indoor localization. In [77], it was demonstrated that FM and Wi‐Fi signals are complementary; that is, their localization errors are independent. Further, experimental results indicated that when FM and Wi‐Fi signals were combined to generate fingerprints, the localization accuracy increased by 11% (without accounting for temporal variation) or up to 83% (when accounting for wireless signal temporal variation) compared to when Wi‐Fi RSSI only is used as a signature. A few efforts have also proposed Bluetooth RSSI‐based localization. Gimbal [67] and iBeacon [68] allow users to instrument their environment with custom Bluetooth‐based beacons and then use RSSI values received from these beacons on the user’s mobile device for indoor localization. Zigbee RSSI‐based localization has also been proposed in several works [2, 78, 79].

Figure 37.5 Measured accuracy and Wi‐Fi signal distributions excerpted for an indoor location. Different locations are often covered by different Wi‐Fi access points (APs) and or have different signal strength characteristics for the same Wi‐Fi AP (as can be seen from the two plots in the figure), allowing for unique fingerprinting. Black, blue, violet, and red bars on the map represent 3‐, 6‐, 9‐, and more than 12 m error distances, respectively, when using Wi‐Fi RSSI fingerprinting for localization [64].

Source: Reproduced with permission of IEEE.

Although empirical RSSI‐based localization schemes are very popular, a disadvantage of these methods, as mentioned earlier, is that they require labor‐intensive surveying of the environment to generate radio maps. Crowdsourcing is one possible solution to simplify the radio map generation process, by utilizing data from multiple users carrying smartphones [80]. Utilizing principles from Simultaneous Localization and Mapping (SLAM) approaches proposed for robot navigation in a priori unknown environments can also be beneficial for quickly building maps of indoor locales [81]. Another challenge is that RSSI readings are susceptible to wireless multipath interference as well as shadowing or occlusions created by walls, windows, or even the human body; thus, in dynamic environments such as a shopping mall with moving crowds, the performance of fingerprinting can degrade dramatically. To overcome multipath interference, a recent effort [82] proposes using the energy of the direct path (EDP), and ignoring the multipath reflections between the mobile client and APs. EDP can improve performance over RSSI because RSSI includes the energy carried by multipath reflections which travel longer distances than the actual distance between the client and the APs.

Several approaches have proposed computer‐vision‐based fingerprinting [83–86]. These techniques require the mobile subject to either carry a camera or use a camera embedded in a handheld device, such as a smartphone. When the subject moves around in an indoor environment, the camera captures images (visual fingerprints) of the environment, and then determines the subject’s position and orientation by matching the images against a database of images with known location. One challenge with such an approach is the high storage capacity required for storing the images of an indoor environment. Significant computing power may be required to perform the image matching, which may be challenging to implement on mobile devices. If subjects are required to carry supporting computing equipment [85], it may impede their mobility. In [87, 88], it is assumed that the illumination intensity and ambient color vary from room to room and thus can be used as fingerprints for room‐level localization. Light matching [89] utilizes the position, orientation, and shape information of various indoor luminaires, and models the illumination intensity using an inverse‐square law. To distinguish different luminaires, the work relies on asymmetries/irregularities of luminaire placement. Absolute and relative intensity measurements for localization have been proposed in [90, 91], respectively, with knowledge of the receiver orientation assumed to be available to solve for a position. These techniques employ received intensity measurements to extract position information from multiple transmitters using a suitable channel model. Other approaches exploit the directionality of free space optics, where angular information is encoded by discrete emitters (e.g. modulated LED‐based beacons) [92–94]. The Xbox Kinect [95] uses continuously projected infrared (IR) structured light to fingerprint and capture 3D scene information with an infrared camera. The 3D structure can be computed from the distortion of a pseudorandom pattern of structured IR light dots. People can be tracked simultaneously up to a distance of 3.5 m at a frame rate of 30 Hz. An accuracy of 1 cm at 2 m distance has been reported.

Some approaches utilize dedicated coded markers or tags in the environment for visual fingerprinting, to help with localization [96–100]. Such approaches can overcome the limitations of traditional vision‐based localization systems that rely entirely on natural features in images which often lack of robustness, for example, under conditions with varying illumination. Common types of markers include concentric rings, barcodes, or patterns consisting of colored dots. Such markers greatly simplify the automatic detection of corresponding points, allow determination of the system scale, and enable distinguishing between objects by using a unique code for different types of objects. An optical navigation system for forklift trucks in warehouses was proposed in [96]. Coded reference markers were deployed on ceilings along various routes. On the roof of each forklift, an optical sensor took images that were forwarded to a centralized server for processing. Another low‐cost indoor localization system was proposed in [97] that made use of phone cameras and bar‐coded markers in the environment. The markers were placed on walls, posters, and other objects. If an image of these markers was captured, the pose of the device could be determined with an accuracy of “a few centimeters.” Additional location‐based information (e.g. about the next meeting in the room) could also be displayed.

An alternative approach to physically deploying markers is to project reference points or patterns onto the environment. In contrast to systems relying only on natural image features, the detection of projected patterns is facilitated by the distinct color, shape, and brightness of the projected features. For example, in [101], the TrackSense system was proposed, consisting of a projector and a simple webcam. A grid pattern is projected onto plain walls in the camera’s field of view. Using an edge detection algorithm, the lines and intersection points are determined. By the principle of triangulation – analogous to stereo vision – the distance and orientation to each point relative to the camera are computed. With a sufficient number of points, TrackSense is able determine the camera’s orientation relative to fixed large planes, such as walls and ceilings.