GB2619703A

GB2619703A - A system and method for crowd density estimation for effective navigation and position improvement

Info

Publication number: GB2619703A
Application number: GB2208365.3A
Authority: GB
Inventors: Nonyelu Fredi
Original assignee: BRITEYELLOW Ltd
Current assignee: BRITEYELLOW Ltd
Priority date: 2022-06-08
Filing date: 2022-06-08
Publication date: 2023-12-20
Also published as: GB202208365D0

Abstract

A system for measuring or predicting the density of a crowd in an environment comprises a plurality of anchor points 301, 302, 303, 305, 306, 307 for collecting received signal strength (RSS) data in a three-dimensional space relative to a portable user device 304. A processor is configured to arrange the RSS data in descending order and to select the strongest anchor points for distance estimation, and a machine learning model is implemented to predict a path loss exponent. An exponent is selected from a range predicted by the model, and a distance value of the portable user device is estimated. The distance value is sent to a gateway 308, which is configured to perform trilateration to calculate the position of the device. The process may be repeated for multiple user devices and the same anchor points to derive a crowd density. The environment may be an indoor environment.

Description

A SYSTEM AND METHOD FOR CROWD DENSITY ESTIMATION FOR EFFECTIVE NAVIGATION AND POSITION IMPROVEMENT

TECHNICAL FIELD

The invention relates to a method and implementing system for the approximation of crowd density (mobile as well as static) in an indoor environment. The invention seeks to remove or ameliorate the duplication of wireless devices, which further improves accuracy for estimating the density of a crowd in an indoor environment. Furthermore, a machine learning model as utilised by the invention can be trained with the purpose of improving the position of devices detected by wireless sniffers (i.e. a device built to capture wireless network traffic and analyse it to generate insights into activity in a network at any given time).

BACKGROUND OF THE INVENTION

The accurate estimation of crowd density in an indoor environment typically requires complex procedures. It is common for wireless devices to follow strict security policies, which adds to complexity of the algorithms used in accurately estimating crowd density. The duplication of wireless devices, is yet another issue, which further increases the complexity of density estimation algorithms. In other words, existing systems may result in an additional detection of the same device that would inflate the numbers in the crowd, beyond what those actually present. This is especially the case if a user has multiple devices because the mac address will be the same.

Based on such strict polices, a received signal strength-based location estimation has become a popular and effective method. However, the effectiveness of received signal strength-based methods requires a line of sight between the tag and the anchor. Unfortunately, rapid changes in an indoor environment and the mobility of a tag increases the non-line of sight probability between anchor points and the tag, leading to a substantial reduction in the accuracy of estimating the tag's location.

SUMMARY OF THE INVENTION

The invention seeks to address identified problems and, in general, alleviate the indoor (or outdoor, mixed) coverage problem to result in seamless and higher precision accuracy. The invention may reduce overall system cost, as well as stabilise the inaccuracies of localisation, or at least provide a viable alternative to available systems.

In a broad aspect the invention proposes a system according to claim 1; i.e. a simple and effective implementation of a three stage (module) method for crowd density estimation.

Particularly, the invention is capable of analysing wireless signals generated by using Bluetooth® as well as Wi-Fi® technologies, and then using machine learning techniques to accurately convert the received signals into distance values. The system also enables the minimisation the duplication of wireless devices (detection of the same device more than once), further improving the accuracy of the people density algorithm.

A solution based on a machine learning algorithm is used to further improve the positioning of devices identified by wireless sniffers. Moreover, in a scenario where wireless sniffer devices are able to obtain the MAC (media access control) IDs of sniffed devices, a second module of the method may be completely ignored and the process follows the combination of the first and third method stages to accurately predict a device location and removes duplication.

The invention enables a determination of crowd density of people in a zone/grid (total number of people within) to assess how busy an area in an environment is. If it is known that a zone is overcrowded based on the crowd density of that zone, it is possible for a user to avoid going there, e.g. to find an alternative route to navigate to a destination. Density may be measured in people (or carried devices) per surface area, i.e. square meter or a larger metric such as ten square meters.

In a broad sense the invention relates to the approximation of crowd density (mobile as well as static) in a given environment. It also aimed at removing the duplication of wireless devices, which further improves the accuracy of estimating the density of a crowd in an environment. A particularly unique aspect is the combination of machine learning (e.g. the random forest model) which is fed with data (x,y cords) and predicts where it thinks that the devices should be placed.

As mentioned herein, the purpose of the invention is mainly the approximation of crowd density in an (e.g. indoor environment), removing duplicate warless devices, to reduce the density of a crowd in such an environment and, finally, the use of machine learning to optimize the position of the sniffed devices.

In an alternative broad aspects the invention involves a method of crowd density estimation, comprised of: collecting received signal strength data in a three-dimensional space from a plurality of available anchor points; predicting the location of a tag or portable cellular device carried by a user. In one form the plurality of available anchor points rely on Bluetooth and/or WiFi technologies. The method may further comprise rearranging the received signal data in a descending order and selecting the best three anchor points for distance estimation. The method may further comprise training a supervised learning algorithm based on received signal strength data to predict a path loss exponent to estimate the distance between each anchor point and a tag or portable device, and to estimate x, y, and coordinates.

In a further aspect the invention may be a method for removing duplication of wireless portable devices in a crowd density estimator, comprising: collection of x, y, and z coordinates data from every available wireless detection device; using correlation formula: fbarj(*) = f x 14; and a predefined a threshold value.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 illustrates the overall system architecture and the elements involved in the process of this invention; Figure 2 illustrates the working steps of an algorithm for a method of the invention; Figure 3 illustrates working steps of a further algorithm for a method of the invention; and Figure 4 illustrates further working steps of an algorithm for a method of the invention.

DETAILED DESCRIPTION OF THE INVENTION

The following description presents exemplary embodiments and, together with the drawings, serves to explain principles of the invention. However, the scope of the invention is not intended to be limited to the precise details of the embodiments, since variations will be apparent to a skilled person and are deemed also to be covered by the description. Terms for components and method steps used herein should be given a broad interpretation that also encompasses equivalent functions, devices and features. In some cases, several alternative terms (synonyms) may be provided but such terms are not intended to be exhaustive.

Descriptive terms should also be given the broadest possible interpretation; e.g. the term ''comprising" as used in this specification means "consisting at least in part of" such that interpreting each statement in this specification that includes the term "comprising", further features other than that or those prefaced by the term may also be present. Related terms such as "comprise" and "comprises" are to be interpreted in the same manner.

The description herein refers to embodiments with particular combinations of features, however, it is envisaged that further combinations and cross-combinations of compatible features/steps between embodiments will be possible. Indeed, isolated features may function independently as an invention from other features and not necessarily require implementation as a complete combination.

Figure 1 outlines a system architecture and the elements involved in the overall process associated with the invention. Particularly, wireless sniffer devices 301-307 are configured to monitor the distribution of people in an indoor space. However, it is noteworthy that the invention is also application to other defined spaces such as an amusement park (with a mixture of outdoor and indoor attractions), zoo, stadium/sporting event or any scenario of a venue that people move through.

This is achieved by collecting beacon signals for Wi-Fi® and Bluetooth®. These beacon signals are used to generate an RSSI (Received Signal Strength Indicator) value between the sniffer devices and mobile phones (304) of people in an area, using a path loss formula to convert RSSI to distance, where the path loss exponent is predicted using a machine learning model, e.g. random forest.

Each distance calculated at each sniffer is sent to a gateway 308 which performs a trilateration process to calculate the position of a mobile device. The MAC addresses are rearranged and only the first three octets are used to maintain anonymity. The MAC addresses of the mobile devices are not saved on the system and are discarded as soon as the position is calculated to maintain anonymity.

As mentioned, devices 301, 302 and 303 represent Wi-Fi® tracking devices, while devices 305, 306 and 307 represent Wi-Fi® and Bluetooth® tracking devices. Gateway device 308 collects all ranging data from devices 301, 302, 303, 305, 306 and 307 relative to the mobile device 304 belonging to a person being tracked in the area being monitored, e.g. an indoor venue such as a shopping mall or entertainment facility.

The gateway 308 acts as a central point which connects a cloud server 310 to a wireless counting system. All location data points from anonymous device IDs of people in the venue are sent from the gateway 308 to the cloud 310 via a VPN router 309. A dashboard 311, which controls a setup process, maintains the status of each device in the network.

Configurations/modifications made to the sniffer devices monitoring device 304 and the gateway 308 are performed via VPN through the VPN router 309. Sniffer devices are given a fixed ID which is mapped on the cloud server 310 to determine a fixed position of the sniffers in the venue. Every unique ID sent from the gateway device 308 to the remote server 310 is shown as a unique marker in the portal. Differentiation between each of the devices can be achieved using modified MAC IDs of the mobile devices (304) in the venue.

In a case where Wi-Fi® signals cannot be tracked, the system utilises Wi-Fi® and Bluetooth® capable sniffer devices 305, 306, 307 to track any missed signals. A Bluetooth® signal RSSI is used for a trilateration formula used to determine the position of the mobile device 304.

Figure 2 illustrates the operational steps of a method 100 according to the invention, where it is noteworthy that wireless sniffers are identified as anchor points. Particularly, the method 100 serves a number of purposes, including: * separating wireless sniffer devices (301, 302, 303, 305, 306, 307) based on the respective received signal strength levels; * excluding the wireless sniffer devices which are in a non-line-of-sight relationship with the other sniffed devices; * repeating the process to ensure that at least the three best (i.e. the highest receive signal strength strongest signal) wireless sniffer devices are selected; * utilise random forest to perform accurate mapping of path loss exponent to distance values; * utilise trilateration to estimate x, y and z coordinates of the sniffed device (304). 10 To estimate a suitable value of path loss exponent a and to predict an accurate distance between the wireless sniffers (anchor) and the sniffed devices (tag), the method 100 utilises a supervised machine learning approach, i.e. random forest. A skilled person will be able to implement, e.g. by steps of utilising Python programming language, creating a random forest model, use appropriate data to train the model, finally, test the predictions.

Essentially, a random forest is a classifier based on a family of classifiers h(x10i) ..... h(x10k) ..... h(x10K), and based on a classification tree with parameters Ok that are randomly chosen from a random vector model 0.

For the classificationf(x) (which combines the classifiers ({hk(X)})), the most popular class at input x is voted on by each tree, and the class with the most votes wins. Considering a general scenario, where multiple devices are present to be sniffed, the attributes of the data sets can be classified into the following four types: * The prediction of path loss exponent a.

* Estimation of actual distance values.

* Prediction of the distance values based on the prediction a.

* The strength of the received signal.

The objective is to build a classifier that predicts the value of a. based on the testing y and training x data sets, respectively, and based on the examples D. The x is fed with the actual distance, the received signal strength and the predicted distance data and the y is fed with the a data, which is required to be predicted.

As part of the proposed invention, the bagging technique is used to train the data set. Using bagging, multiple models are fitted to different subsets of a training data set, and predictions from all the models are combined. A mathematical expression of the bagging process is given by the following: Tree learners are trained using the general technique of bootstrap aggregation, or bagging.

Given a training set X = x,, with responses yn, the bagging repeatedly (B times) replaces the training set with a random sample and fits trees to the samples and is composed of the following components: * Sample, with replacement, n training examples from X, Y, call these Xb and Xb..

* Train a classification or regression tree fb, on Xb and Vb.

The prediction of unseen samples x', after training, can be made by averaging the predictions from all the individual regression trees on x' or by taking the majority vote in the case of classification trees. A free parameter for B is the number of samples/trees. A training set typically contains a few hundred to several thousand trees, depending on the size and nature of the data. Cross-validation, or observing the out-of-bag error, can be used to determine the optimal value of B. Of the data set, 80% is used for training and 20% is used for testing. After splitting of data, a random forest algorithm is fit into the training set and RandomForestClassifier class is used. In addition, since the model is fit to the training data set, it is possible to predict the value of a.

A root mean square error matrix is used to evaluate the machine learning algorithm's accuracy, and it is expressed as follows: 21ISE (2) The random forest uses this formula to determine which branch is the best choice based on the distance between each node and the predicted actual value. The yi value represents the number of data points that have been tested at a certain node and theft value represents the value returned by the decision tree. N represents the number of samples/trees.

Once a has been predicted by the random forest method, the following equation (3) is used to estimate the distance value based on non-line of sight received signal strength: d, = 075>< (3) where, alms; is the distance between the wireless sniffers and the sniffed devices, and Prssi is the strength of the received signal.

Figure 3 illustrates the working steps of a second method module 200, initiated if wireless sniffers are unable to get the MAC IDs of sniffed devices 304, e.g. due to security reasons.

The method 200 serves several purposes, e.g. as follows: 1. Uses method 100 to estimate x, y, and z coordinates of sniffed devices (or tags as shown in Figure 3); 2. Estimates the difference in distance values between each of the wireless sniffed devices, i.e. taking a difference of each corresponding coordinates; 3. Calculates a correlation between each of the sniffed devices and, if it is below a present threshold, no duplication is found; otherwise a device is removed from the area being estimated to avoid the same device being detected twice or more; The value of correlation r can be calculated using the following formula: = _ - (4) where xi, yi, and z, are the x, y, and z coordinates of sniffed device, ft, and f are the mean of values available in the data set, being compared. In the present scenario and at any given instant, the data set contains only two values, as at any given instant, two sniffed devices are being compared.

Figure 4 illustrates the working of a third method module 300 initiated if wireless sniffers are able to get the MAC IDs of sniffed devices. The method 300 serves the following purposes: 1. If the MAC IDs are available, method 200 as described above is ignored and the duplication of devices is removed by using the MAC IDs.

2. If wireless sniffers are able to obtain the MAC IDs of sniffed devices, the method 100 is used to analyse the selected received signal strength values for location estimation, to estimate the path loss exponents and use trilateration to find x, y and z coordinates of each sniffed device.

3. Collect the actual x, y and z coordinates of the sniffed devices and preprocess them to the right format.

4. Feed the pre-processed data to the random forest machine learning model; train the model and predict the new x, y and z coordinates based on where the machine learning model believes that the sniffed devices should be located.

A method of the invention also involves further improving the positioning of the sniffed devices from wireless sniffers, by training and then reusing the random forest algorithm.

The method is comprised of training the random forest algorithm on initial x, y, and z coordinates data sets. The random forest is usually used to obtain a better accuracy in predicting the actual location of the sniffed devices; the training should be performed on each x, y, and z coordinates data sets separately.

The method may be comprised of collection of x, y, and 7 coordinates data from every available wireless sniffer, to predict the actual x, y, and z coordinates of the sniffed devices by using the already trained algorithm.

The invention may be summarised as a method and system for determining the location of a wireless mobile device in an environment (e.g. an indoor, outdoor or mixed venue such as a shopping mall, train station, airport, stadium, etc), relative to a plurality of anchor points for the purpose of determining crowd density. A signal strength is measured between the device and an anchor point and the strengths arranged in a descending order for selection of the strongest anchor points for distance estimation. A machine learning model is implemented for path loss exponent prediction; selecting a path loss exponent from a range predicted by the machine learning model; used for estimating a distance value of a portable user device. The distance value is sent to a gateway configured to perform a trilateration process to calculate the position of the mobile device.

Claims

Claims: 1. A system for measuring/predicting the density of a crowd in an indoor environment, comprising: a plurality of anchor points configured for collecting received signal strength data in a three-dimensional space relative to a (plurality of) portable user device(s); a processor configured for arranging the received signal strength data in a descending order and selecting the strongest anchor points for distance estimation; implementing a machine learning model for path loss exponent prediction; selecting a path loss exponent from a range predicted by the machine learning model; estimating a distance value of a portable user device; sending the distance value to a gateway configured to perform a trilateration process to calculate the position of the mobile device.
2. The system of claim 1, wherein an identification address of a mobile device is truncated to maintain anonymity.
3. The system of claim 1 or 2, wherein an identification address of a mobile device is not saved and is discarded as soon as the position is calculated, to maintain anonymity.
4. The system of any preceding claim, wherein position data points from portable user devices are sent from the gateway to a remote server, via a VPN router.
5. The system of claim 4, wherein reconfiguration commands can be sent to the anchor points via the VPN.
6. The system of claim 4 or 5, wherein anchor points are given a fixed ID mapped on the remote server.
7. The system of any preceding claim, wherein the plurality of anchor points utilise Wi-
8. The system of any preceding claim, wherein at least one of the plurality of anchor points utilise both Wi-Fi® and Bluetooth®.
9. The system of any preceding claim, configured for removing duplication of a detected portable user device by the steps of: obtaining identification address of portable user devices; analysing the selected received signal strength values for location estimation; estimating the path loss exponents and using trilateration to find x, y and z coordinates of each detected portable user device; collecting the actual x, y and z coordinates of the portable user devices and preprocessing them into a predetermined format; sending the pre-processed coordinates to the machine learning model; training the model to predict new x, y and z coordinates based on where the machine learning model believes that the portable user devices should be located.
10. The system of any preceding claim, configured for removing duplication of a detected portable user device by the steps of: determining the difference in distance values between each of the detected portable user devices, by taking a difference of each corresponding coordinate; calculating a correlation between each of the detected portable user devices and, if it is below a preset threshold, concluding no duplication is found; otherwise a detected portable user device is removed from the area being estimated to avoid the same device being detected twice or more.
11. The system of any preceding claim, wherein the three strongest anchor points are selected from the received signal strength data.
12. The system of any preceding claim, wherein the machine learning model is a random forest.
13. A method for predicting the density of a crowd in an environment, comprising: a) collecting received signal strength data in a three-dimensional space from a plurality of anchor points relative to a portable user device; b) arranging the received signal strength data in a descending order and selecting the strongest anchor points for distance estimation; c) implementing a machine learning model for path loss exponent prediction; d) selecting a path loss exponent from a range predicted by the machine learning model; e) estimating a distance value of the portable user device repeating steps a) to e) to estimate distance values between multiple user devices (tags) and the same anchor points to derive crowd density.