------------------------ Submission 2619, Review 2 ------------------------ Title: FormaTrack: Tracking People based on Body Shape Primary Associate Editor's recommendation to Editors Acceptable with minor (or no) changes by IMWUT Primary Associate Editor meta-review The paper has received some quite differing scores. However the substance of all reviews was quite similar. The general idea and some basic results are very interesting. However there were some suggestions/claims with respect to scalability to larger buildings and more people which are just too speculative. Also, the reviewers requested detailed information about the power consumptions. In this sense the minor review as such has to be taken seriously and will be carefully evaluated by the AE Primary Associate Editor Meta-review: Major / minor revision requirements - Make sure that claims related to building level tracking and more people are scaled back. - Provide absolute power consumption numbers Additional comments for authors (blank) ------------------------ Submission 2619, Review 1 ------------------------ Title: FormaTrack: Tracking People based on Body Shape Contribution to IMWUT The authors have taken a single-chip ultra wide band impulse radar transceiver which has been used for sleep, respiratory monitoring and occupancy and with the aid of supervised machine learning they have produced a system to differentiate people moving through doorways in different ways. In effect a person detection system, at a room level. By taking all the distance data from the radar signal they can improve over height measures alone (~80% to 90%) in a 2 person in a group configuration. When considering a group of 8 with heights from 161-183cm and 60.8kg to 97.9kg their accuracy is reduced to a little over 60%. However, the can demonstrate convincing results when larger room configurations are considered. Review The authors have used a XeThru radar which affords a single chip (monostatic) implementation of an Impulse Radar system. The company which produces this offers various distance, presence and motion features in their products. The authors are interested in proposing a new biometric feature called body shape. However, this is not well defined one introduced. Section 1 could be improved with a diagram which illustrates what areas of the body are considered part of the body shape metric. The later discussion and results show that height is a large determining factor and perhaps then head to shoulders in the next largest factor. Does the actual shape of the torso, arms, hands, face and legs actually have anything to contribute here? Instead is it simply head distance and shoulder distance and cross section of head and shoulders? More detail should be provided on how you have emulated 15 different floors plans. Currently it suggest an arbitrary selection of floor plans and random movements were used, whereas you would not expect some configurations to result in all possible movement patterns being seen. I would like to see test results of just straight movement through the doorways being reported on in terms of accuracy. I fear the random movements + arbitrary floor plans are resulting in the simulation of more unique movement pattern that might be expected in a real deployment. For a journal paper the related work section is quite brief. Some related work in tracking or tracking objects or use of Radar and ML with recognizing body parts Powerline positioning: A practical sub-room-level indoor location system for domestic use Recognizing daily activities with RFID-based sensors RadarCat: Radar Categorization for Input & Interaction A survey of indoor positioning and object locating systems Object tracking: A survey There are a wealth of location based systems, room level tracking and presence detection research and use of Radar for sensing objects and body parts. I would like to have seen a more systematic treatment of description of how these pieces of related work were selected (e.g. reporting of a systematic literature review process) There are about 5 pages of radar and measurement description provided here. What is not clear is how much this contributes to the paper. For example, the XeThru radar provides a distance measure but this isn't considered or compared with in section 3.2 for distance estimation. Section 3.3 is very limited. I was expecting to be presented with details on what is happening as a body moves through the doorway in terms of the reflected signal at key points. As noted earlier the use of blackbox for your machine learning means you cannot detail what aspects of the body are contributing to the reflected signal and hence the metric. The experimental setup and evaluation are well described and well reported. Recommendation(s) to 1AE Acceptable with minor (or no) changes by IMWUT Major / minor revisions (recommendation to 1AE) Improvements on the related work Diagrams, as noted, to improve the presentation of the work More details in section 3.3 (e.g. aided by diagrams) Confidence in the review Very confident - I am knowledgeable in the area ------------------------ Submission 2619, Review 3 ------------------------ Title: FormaTrack: Tracking People based on Body Shape Contribution to IMWUT This paper describes an ambient system that recognizes identity and the direction of a single person when he/she is passing through a door. The system doesn't raise privacy concern as camera and microphone. The evaluation demonstrates a quite attractive accuracy. Review In general, this is a really nice paper that I enjoyed reading. + a clever idea: the system setup contains sufficient information for identifying a person, even when the person is carrying different extra things (e.g. cap, mobile phone), and on the other hand doesn't capture too rich information that might raise privacy concern. + the paper is well-organized and well-written (e.g. structure, language). + the experiment is well-designed considering multiple aspects like different walking path, the difference between day to day. The number of participants is also OK (except one experiment) + multiple aspects of the system are considered, including power consumption, the time needed for training. These are essential for real implementation. There are only some small points that need further clarification and improvement. . in the abstract, it's declared, that "we expect that other, more sophisticated radar sensors can achieve high accuracy with more people and in larger buildings". I feel it's a little bit exaggerating, I guess with more sophisticated radar sensors, the algorithms also need to be changed. How many percent of content in this paper can still be used remains unclear. I'd rather either not claim it or provide more details (what kinds of radar sensor, what further algorithms, what kind of improvement etc.). . the participant number in the last experiment is only 2. This is pretty small. I fully understand the difficulty in data gather and processing, but with a number of 2, the result (although very promising now) is not very convincing. . it's very nice to consider the power consumption and it's great to see power consumption can be dropped to 30%. However, it would be nice to provide the power consumption in number (viz., in Watt), and considering how much power is used in data gathering and how much in data processing . If more data can be provided on other typical door passing system (how much power they consume) so that the reader can compare between the FormaTrack and them, it will be very welcome. Recommendation(s) to 1AE Acceptable with minor (or no) changes by IMWUT Major / minor revisions (recommendation to 1AE) I'd highly suggest increasing the number of participants to at least 5 (if not already more than 5) in all experiment. Confidence in the review Very confident - I am knowledgeable in the area ------------------------ Submission 2619, Review 4 ------------------------ Title: FormaTrack: Tracking People based on Body Shape Contribution to IMWUT The article is about a new approach to detect and identify persons passung through doors. The underlying technology utilizes radar information up to 1m. The authors discuss after their introduction and problem definition related work and after that present their idea utilizing radar information. They therefore give a short introduction into the technic of radar based sensing including the possiblities of relying on the so called doppler effect. The step by step improve the system to reduce false positives and introduce new algorithms for determing the direction from which side the person is entering the room. Applying machine learning approaches the authors can differentiate between two persons with close to 100 percent confidence which drops to 80 person if they try to differentiate 4 persons. In a simulation they surveile the possibility to track up to 4 persons in different floorplans. In another chapter the authors discuss the power reduction by either combining the system with a pir sensor or by reducing the sampling rates of the system. Review Although I really like the idea of this paper using radar sensors to detect the movements of persons throughout the houshold I have mixed fealings about this work: 1) Structure: Initially the authors discuss their idea of using radar sensors to detect the movements introducing an abstract description of the problem. Coming from data processing and anaylis, I would have prefered learning more about the used radar systems at the beginning of the article (which would have avoided many thoughts untill I reached the evaluation part). Issues which I would have adressed earlier: Sampling rate of the used system, range 2) Discussion of the power consumption: There are no number about the power consumption of the system without power saving features. 3) What happens if a very large person walks through the door so that the head is closer to the sensor than the given noise threshold? 4) Health considerations with regard to large persons 5) Tracking of persons throughout the house - These are only hypothesis - The simulation gives you hints - but from a single door to an complete house - there are more issues which have to be considered... (And this chapter gives the reader the feeling of reading an advertisement for your other papers ;)). 6) Evalutation / Testing: The authors record the weigh of persons only once (or perhaps they ask them), RF Reflections / signal attenuation strongly depend on the fraction of water being part of the obstacle, therefore I think the article should also consider this. In the training / evaluation using machine learning approaches, the authors can achieve this by using random samples of a data set from the persons from each day throughout the week (which would consider the changes in water ratios - especially if subjects come from gym...) Recommendation(s) to 1AE Reject Major / minor revisions (recommendation to 1AE) (blank) Confidence in the review Highly confident - I consider myself an expert in the area ------------------------ Submission 2619, Review 5 ------------------------ Title: FormaTrack: Tracking People based on Body Shape Contribution to IMWUT This paper presents a study on the use of a body-shape detecting, monostatic radar (mounted on a door frame) to track people in a multi-occupancy home. The paper includes an in-depth description of a method for 1) detecting when someone walks through a doorway, 2) determining direction of travel, 3) identifying the person from a group of up to 8 people. The authors further examine the use of this technique by incorporating a simulation of a multi-room, multi-person scenario. The main study is carried out using 8 people over 7 days, and includes variations in clothing. In an additional study, 2 people are analysed while carrying different objects. Overall the methods described are of sufficient detail for reproducibility, and the evaluation makes a convincing case for using this method in future work. This would be an excellent contribution to IMWUT, of interest both to those working on home sensing and sensing more generally. Review I like this work. Almost all of my potential comments were addressed as the paper went on. Those that remain are detailed in the 'Minor Revisions' section below. The biggest concern I have is with my (mis) interpretation of Figure 3 (see below). In brief, why is the non-crossing (noise) distance not higher than the the ends of the crossing distance? Some additional thoughts while reading: It is interesting that high-heels can greatly affect the results (why is this result not in Fig 16?) It would seem that the ideal training might involve instances of each person wearing their most variously shaped clothing (heels, sandles, big hats, etc.) The addition of extra sensors to save power makes sense, although it does have the limitation of extra installation costs. [I'm curious if the information from those sensors might even be added to the radar to improve the accuracy of the direction algorithm.] I would also be interested to know what the actual power usage of the equipment is. Recommendation(s) to 1AE Acceptable with minor (or no) changes by IMWUT Major / minor revisions (recommendation to 1AE) I have a number of issues that I would like to see addressed: 1. section 3 onwards refers to 'non crossing', could you briefly define exactly what this means early on (i.e. no-one is there, rather than someone standing still under the door) 2. I am confused by Fig3 (a) and the accompanying text in 3.1. Why would the 'non-crossing' distance be lower than the 'crossing' event? Judging by the response of a crossing event, I would have expected the non-crossing distance to be closer to 1.0. Is this some kind of distance range measure rather than an absolute distance? Could you please explain this? 3. Also with Fig3 (b), why are all of the non-crossing events within the first 40 or so events? This figure needs to be better explained. 4. In 3.2 the "Radar Frame Rate" is mentioned but is nowhere defined. Later we learn that the "frame rate" is 170Hz, but how does this fit with the formula in 3.2 ( 170/(2*K) Hz doesn't seem right...)? Or is F some other number. Please clarify this and the text following equation 3. 5. Further limitations. Out of curiosity: what happens if a person stops and waits for a period of time on one side of the door? Similarly if a person changes speed while walking through the door. 6. If available, what is the actual power consumption of the radar device? Minor comments: - on writing: 3.2, p9. "For this we tilt FormaTrack towards one of the rooms" -- this is written here as if its the first time the idea was mentioned, but the first paragraph of 3.2 already introduces it. Perhaps re-phrase here as a reminder. - style: the end of p9 looks like an unfinished sentence (because the eqn. falls onto the next page). Personally I prefer to see some form of punctuation indicating what is to follow, e.g. a colon. - 4, p10 "i.e. once each day" -- should that not be 25 times each day? Confidence in the review Highly confident - I consider myself an expert in the area