Going My Way: a user-aware route planner

Transcription

1 Going My Way: a user-aware route planner Jaewoo Chung Media Laboratory, MIT 20 Ames St. E15-384C Cambridge, MA USA jaewoo@media.mit.edu Paulina Modlitba Media Laboratory, MIT 20 Ames St. E15-384C Cambridge, MA USA paulina@media.mit.edu Chaochi Chang Media Laboratory, MIT 20 Ames St. E15-384C Cambridge, MA USA ccchang@media.mit.edu Abstract Going My Way is a mobile user-aware route planner. The system learns a user s everyday routes and provides directions from locations along those routes. The mobile phone client application logs GPS information in real-time, and translates this information into a route model. When a user requests directions to a destination, the phone client sends the route information to our custom server application, which then retrieves the directions from on the basis on the user s every-day route to the desired destination. Going My Way provides directions, if available, based on personal landmarks rather than street names and intersections. The main goal is to reduce the user s cognitive load by simplifying and personalizing directions; guiding the user to his or her destination by using knowledge of where the user has been and what he or she cares about. Keywords Personal navigation, pedestrian navigation, location awareness, personal landmarks, mobile computing, human-computer interaction, HCI, context awareness Copyright is held by the author/owner(s). CHI 2008, April 5 April 10, 2008, Florence, Italy ACM 1-xxxxxxxxxxxxxxxxxx. ACM Classification Keywords H5.m. Information interfaces and presentation (e.g., HCI): Miscellaneous.

2 2 Introduction Consider a situation in which you ask a friend of yours for directions, for example to the restaurant Kaya in Cambridge, MA. Rather than describing the whole route, your friend probably would begin by asking you about other places, located near or on the way to the destination, which you may be familiar with. These places may be public landmarks or just locations (we call them personal landmarks ) that you and your friend have visited together. Alternatively, your friend may know you well enough to feel comfortable with guessing which places you are familiar with. By using the knowledge, your friend then provides you with directions from that personal landmark to the destination: You know that store on Main Street that sells funny T-shirts? Restaurant Gaia is just across the street from it. On the other hand, the directions that you get from route planning systems and applications, such as web based map services (e.g. Google, Yahoo) or car navigation devices, is normally not based on knowledge about which locations are familiar to you. Some map and navigation systems allow users to mark waypoints as intermediate stops along the destination. This option can be used to reroute the direction to include the user s familiar paths, but the option requires that the user makes the effort to manually manipulate the direction based on his or her recognition of locations on the map. In the other hand, MyRoute [12] is able to generate directions based on a user s familiar locations that are close to the destination. The main limitation of this approach is that the user need manually provided the user s familiar landmarks in order to personalize the directions. Especially in cities, where the number of landmarks is large, manual systems quickly become inefficient since the users would have to enter a large number of locations in order to maximize the system s usefulness. Thus, automatically detecting and tracking the users locations is inevitable when it comes to solving these limitations. Today, it is possible to detect a user s salient locations by using various location-based techniques, such as clustering algorithms [1][4][10]and tracking of GPS signal loss in indoor locations. [9] However, these techniques cover only a limited number of the places that the user may recognize, for example home and work but no locations in between. Although the detected salient (i.e. often visited) locations can be used as landmarks, a user may also recognize other locations and buildings along his or her frequently visited paths (e.g. the Post Office or Starbucks), although he or she never actually visited the specific location. People often use these landmarks to navigate from one place to another and use these landmarks to give people directions. [8][3] This approach is useful when it comes to finding a new, unknown location in a familiar territory, since it is likely that many landmarks within that territory are familiar to the user. Thus, the sought-after destination could be around the corner from the user s local grocery store, or adjacent to the street the user walks from the subway train to work. Therefore, in our system, we pay more attention to the information along the paths than to the endpoints and salient locations of the user.

3 3 In this paper, we present a system, Going My Way, which aims to detect and utilize information about the user s personal landmarks, which are recognized along his or her frequently visited paths, in order to guide the user to his or her final destination. The remainder of this paper contains a more detailed description of the system and interface, as well as of the main user study that underlies the system s personal landmark algorithm. Approach As described in the introduction, the main goal of the Going my way system is to improve route finding systems by implementing more human-like directions. In order to achieve this, three steps are required. The steps are as follows: (1) collecting the user s location information to enable the system to identify the user s traveling patterns, (2) identifying personal landmarks that are as close as possible to the desired destination, given by the user, and (3) generating direction. The following subsections will describe each of these three steps in depth. Collecting the GPS trace for identifying frequent path: A GPS equipped mobile device, such as a cellular phone or a car navigation system, is required to log location information and enable our system to acquire information about where the user has been. The accumulated GPS data is then used to generate the user-specific route model that contains the user s frequently visited locations and paths. The route model in Going My Way consists of two layered squires, high and low resolution grid systems, sized 50 meters and 1.6 km. Each cell in the high resolution grid has the property of the number of hits, newly collected GPS geo-coordinates, speed and accuracy. Each cell in the low resolution grid contains the corresponding high resolution cells that are covered by it. When a new GPS coordinate is received by the device, the system registers the coordinates to corresponding high resolution grid and increments the number of hits in the cell. If the cell s GPS coordinateaccuracy is higher than the newly received one, the system does not update the coordinates of newly obtained one to keep the model in higher accuracy. When the user travels between locations in his or her daily life, the number of hits increases. This model naturally captures both the significant places and frequently visited path. When the user asks for directions to a destination, the high hit number cells are used for select landmarks around/near the destination. Preparing personal landmarks that are close to the destination: Personal landmarks are generated automatically by the system when the user request for directions to a desired location. The destination can be provided as an address (e.g. 95 Main Street), a service description (e.g. the Post Office), or a specific company or location name (e.g. restaurant Kaya). When the destination information is submitted, the system uses the Geographic Information System (GIS) to get the GPS coordinates of the destination, and thereafter identifies the low resolution cell (proximity 2.5 km 2 ) that covers the location. Within the cell, the system attempts to select the 10 cells that have the highest hit numbers. Then, the system searches for another 10 cells in 1 st peers (adjacent 8 cells which cover proximately 20 km 2 ). These cells are negatively weighted based on the distance from the destination

4 4 and positively based on the number of hits. Finally, the system picks the 10 mostly weighted cells as landmarks. The system avoids picking landmarks from two adjacent areas by checking the distance between the landmarks. The main problem with this approach is that we do not know whether these locations are on or near the landmarks that the user actually recognizes. When the system picks landmarks, it can pick landmarks from a cell that contains identified salient locations. However, the system may also need to pick landmarks from cells that only contain paths between salient locations. For instance, a user may pass by a specific Starbucks coffee shop every day but never actually visits the shop. The system needs to pick locations based only on the fact that GPS traces were collected nearby the coffee shop. Should the system randomly select the location? Why is that particular Starbucks a better selection than a restaurant nearby? One way of picking landmarks is to implement user preference profiles. Another way is to find a general recognition model of the user s recollection of locations. We have chosen the latter approach. In the next section, we present the results of an experiment on which our user model has been built. User Interface: Getting a list of directions requires only a few steps. First, the user needs to provide an address or the name of the destination in text format by using the phone keypad. When the system finds more than one location for the submitted address (or the place name) the system lists the found locations and asks the user to selecting an item from the list. Figure 1. Left picture shows the snapshot of destination input screen, and right picture shows the disambiguating screen When the target location has been identified, the system shows a list of computed personal landmarks that are (1) close to the target location and (2) are likely to be recognized by the user. Landmarks are provided as text descriptions of the location (the exact name is included if possible), e.g. Starbucks in Central Square, and the linear distance from the landmark to the destination. The information format was chosen based on the results of our main experiment, presented below. Generating Direction: Directions from the confirmed landmark and the desired destination is then generated in text format. If a landmark has a specific name, e.g. Star Market, the name is included in addition to the specific address.

5 5 personalized directions. The system version described in the paper collects location and route information automatically in order to provide personalized landmarks. Implementation: The system contains of two main parts: a GIS server (back-end) and a phone application (front-end). The phone application was developed using Java for Micro-Edition (J2ME) on a Motorola iden 870 phone, and the server was developed on the C#.NET platform for Microsoft Windows XP. Figure 2. A screen shot of a phone showing directions. In the previous prototype of Going My Way, we let users label their salient location manually. In addition, the users could enable the system to record and learn routes between those labeled locations. The userprovided contextual information - labels and users frequent route information between them - allows the system to generate more natural directions based on the user s own experiences. Below, follows an example of a learned route that connects a user s home and office with a route via Arlington Street: On your way HOME from the OFFICE, turn LFET at the Starbucks (on Arlington St.) onto Medford St. for 60 meters. Arrive at the post office. Because the system knows the current location of the user, and is able to identify the known location closest to the final destination, the system is able to generate The phone application, in its turn, also consists of two parts: the route learning algorithm and the user interface (for fetching and showing directions). The route learning algorithm was developed based on a previously developed system called Contella [2]. The main function of the interface is to pass text information (an address of the destination) to the server and display the directions thereafter returned by the server. Our GIS server is built on top of Microsoft s MapPoint API. The server finds the nearby names of restaurants and hotels and generates a list of directions. The server and the clients communicate via UDP over an iden data networks. Experiment Setup As a part of the system design and implementation, we conducted a set of experiments to study how people, in general, recognize and memorize different types of objects (e.g. business, buildings, signs and monuments) at different types of locations. We also chose to study if the way in which the location information is presented (text description, address, or

6 6 image) influences people s perception of them. We started with the hypothesis that: 1. People recognize objects which are located at intersections better than objects that are located somewhere along a street. 2. People recognize and locate well-known chains (e.g. Starbucks) better than unique places and stores. 3. People recognize and locate locations better when they are presented as descriptions in text (e.g. Starbucks right next to the big Star Market ) than when they are presented as addresses (e.g. 95 Main Street) or as images. 4. The more time people spend in an area, the better they become at recognizing and locating buildings and locations in that particular area. Experiment settings: We recruited 12 subjects for the experiment; 6 women and 6 men of various nationalities. Apart from one subject who is currently working at MIT, the subjects are all graduate students at MIT. The subjects all mentioned either walking or biking as their main transportation mode. Half of the subjects were new to the area and had lived there no more than 2 weeks; the other half has been living in the area for more than a year. In the first phase of the experiment, we gave the participants a simplified map of the area and asked them to mark the streets that they have visited at least once with a blue marker. Then, we asked them to mark the streets that they are most familiar with with a red marker. Based on the results, we identified the streets that all subjects claimed to be familiar with an overall distance of approximately 2.5km. Along these streets, we then selected a set of 20 locations that are either at an intersection (10) or somewhere along a street (10), and that are either part of a chain (11) or are unique to that area (9). In the second phase, we divided the participants into three subgroups (Group A to C) and presented each location in one of three possible ways (text, address, image), as seen in Table 1. Each group consists of two newbie and two residents who lived the place for more than a year. Group A Group B Group C Place 1 Image Place name Address Place 2 Address Image Place 20 Place name Address Place name Image Table 1. The table shows how the representation types of the locations are distributed to each group. The participants were asked to fill in their answers in an electronic questionnaire, as shown in Figure 3.

7 7 After the experiment, the answers were examined and compared to a pre-marked key map to determine whether the subjects actually recognize the objects and locations and remember their correct location. During the correction procedure, an error margin of 1 block was applied for locations and buildings that are located along a street. For buildings and locations at intersections, the subjects had to identify the correct intersection for it to count as a correct answer.. Results To conclude, the study results both confirmed and contradicted our hypotheses. In this section the results that we think are most significant are described and discussed. The results are directly compared with the hypotheses stated above (see section Experiment Setup). Figure 3. The screen shot of the experiment questionnaire. First, the subjects were asked to find whether they recognize the presented objects or locations at all. As mentioned above, the objects and locations were presented as either a text description (e.g. Starbucks by the Main Street Subway station, an address (e.g. 95 Main Street ), or an image. No cross-testing (e.g. image and text simultaneously) was conducted in this particular experiment. If the subjects recognized the location, they were asked to mark out the location on a map. Then, the subjects were asked to specify if the location is at an intersection or somewhere along a street, as well as specify the full address of the location (if they know it). Finally, subjects were asked to describe what else, if anything, is near the location. In the user study, the overall uncorrected recognition rate was 130/240=54.2%. Out of these 130 answers, 57 (43.8%) were incorrect (false positive/error I). 1. People recognize objects which are located at intersections better than objects that are located somewhere along a street. Result: True. Out of a total of 130 locations and buildings that were marked as recognized, the subjects claimed to recognize 63 (48.5%) as located in intersections. 56 of these claims were correct (Group A: 21; Group B: 35). Thus, the accuracy rate was 56/63=88.9%. 11 answers were not sure. When it comes to locations and buildings that are located along a street, the subjects thought they recognized 42 (32.3%), but 15 of these were incorrect. Thus,

8 8 a total of 27 (A: 9; B: 18) were correctly identified; an accuracy rate of 27/42=64.3%). The remaining 13 answers were not sure. 2. People recognize and locate well-known chains (e.g. Starbucks) better than unique places and stores. Result: False. Whereas a total of 29 out of 132 chain/franchise stores and restaurants were correctly recognized (A: 12; B: 17), 47 out of 108 unique stores and restaurants were correctly identified (A: 20; B: 27). Thus, the overall recognition rate is 29/132=22.0% (accuracy rate: 29/55=52.7%) for chains and 47/108=43.8% (accuracy rate: 47/74=63.5%) for unique places and buildings. 3. People recognize and locate locations better when they are presented as text descriptions (e.g. Starbucks right next to the big Star Market ) than when they are presented as addresses (e.g. 95 Main Street) or as images. Result: True. Totally, 52 images were marked as recognized, of which 26 were incorrect and 26 correct (recognition rate: 26/80=32.5%; accuracy rate: 26/52=50%). The corresponding numbers for text description and address are 44/14 (recognition rate: 30/80=37.5%; accuracy rate: 30/44=68.2%), and 34/17 (recognition rate: 17/80=21.3%; accuracy rate: 17/34=50%). Thus, although a larger number of subjects claimed that they recognized a building or a location when presented with an image, the error rate for images was higher than for both text descriptions and addresses. 4. The more time people spend in an area, the better they become at recognizing and locating buildings and locations in that particular area. Result: True. A total of 240 questions were asked during the experiment; 20 per subject. Thus, out of these 240 questions, 120 questions were answered by subjects who were new to the area, and 120 questions were answered by subjects who were familiar with the area 1. The overall recognition rate for the subjects was (130-57)/240=30.4%; (51-24)/120=22.5% 2 in group A and (79-33)/120=38.3% in group B. Thus, the overall error rate for the two groups was (A) 24/51=47.1% and (B) 33/79=41.8%. When it comes to accuracy rate (correctly recognized locations/locations perceived as recognized by the subject), group A got an equally good or better rate than group B for text descriptions (A: 69.2%; B: 66.7%), addresses (A: 50%; B: 50%), places/buildings in streets (A: 64.3%; B: 64.3%), and chains (A: 63.2%; B: 47.2%). This could indicate that (1) text descriptions are best for both people who are familiar with the area and for people who are relatively new 1 From now on we will refer to these two sub groups as Group A (unfamiliar with area) and Group B (familiar with area). 2 Here 51 is the number of times the subjects claimed that they recognize the location/address/building, 24 is the number of errors among those 51, and 120 is the total number of occurrences.

9 9 to the area; (2) people who are new to an area register and memorize well-known chains, whereas people who are familiar with an area register unique stores and restaurants. However, these theories require further studying and more robust proof. Number of correctly recognized buildings and places personal profile, then, when requested, the system automatically generates directions based on the provided these landmarks. However this approach requires the user s manual effort to provide salient locations into the system and do not take account of information that users may picked up during the journey on the routes between the user provided locations. Number of correct answers Category intersection street unique chain image text address Figure 4. Chart showing the results across the categories. To conclude, our results suggest that rich text descriptions of unique and original places and buildings, located at intersections, are most reliable when it comes to personalized mono-modal directions. The study shows that these descriptions have both the highest recognition rate and the lowest error rate. Related work The work by Patel et al. (2006) [12], MyRoute, takes similar approach that this paper is describing, that is, providing directions based on a user s familiar locations that are close to the destination. MyRoute lets the users to manually save their salient locations in a Several techniques for detecting users salient locations have been introduced by many works. Our previous work by Marmasse and Schmadt (2000) [9] used GPS signal loss and corresponding time-elapse to detect the length of a user s staying at indoor to identify the user s most salient everyday locations. Clustering algorithm is one of the most popular methods for finding people s salient locations using GPS and WiFi hotspots. Kang et al. (2004) [4] detect WiFi hotspots and use time-distance based clustering algorithm to identify the boundary of the significant places. Ashbrook and Starner (2003) demonstrated that and hierarchical clustering algorithm combining with GPS dropout are efficient to find salient indoor and outdoor locations. Liao et al. (2005) [7] showed that not only detecting salient locations but also few types of activity on that location using the previous work of [1] and combining the Relational Markov Networks. Other researches paid more attention on frequent paths between salient locations. Our previous work by Marmasse (2004) [11] collects multiple traces of routes between two significant places and generates template that estimates the frequent route between the two locations. However, the estimate does not necessarily represent and models the actual streets of the frequent

10 10 path that is hard to be used to infer the nearby landmarks. Our later work by Chung (2006) [2] developed Contella that models streets of a user s frequent paths between locations. This model is sufficient enough to be used to extract nearby landmarks that the user may experienced and learned while travel on the route. The limitation of the system is that the user needs to train the system in order for the system to learn routes. The work by Liao et al. (2004) [6] developed the system learns and infers transportation routines such as frequent paths, decision-making points for switching transportation modes (i.e. bus stops, parking lots.) The work by Krumm (2006) [6] created Open World Model that uses probabilistic model that measure the likelihood of being in the 1km sized grid. The model combines user s specific history of transit pattern to increase the prediction of the destination. This grid model is similar to our grid system that computes the likelihood of entering a cell of the grid. However, the Open World Model s cell is too large to be used to extract nearby landmarks as reference points to the destination. Conclusion and Discussion In this paper we have presented a novel mobile route planner. The main contribution of this system is that it shifts the focus from general salient locations to the user s own navigation and exploration experiences. The user-specific information enables interactions that are richer, more usable, and simple than the interactions supported by current navigation and route planning interfaces. Among other things, our studies show that although some basic conclusions can be drawn regarding the way people navigate, the task of personalizing directions is very complicated. The way we navigate is very personal. For example, one subject said: I noticed that little unique restaurant because we had a funny store with the same name in my home town. Still, our studies indicate a number of factors that seem to make places and building more recognizable to people in general; unique/original buildings and urban objects that are located at an intersection and are described with a rich text, such as the hospital right next to the downtown mall are recognized more often and more accurately than other types of urban objects and representation. Future work In the near future, more targeted and complex user studies will be conducted with mobile phones, in reallife, in order to explore some of our old and new hypotheses further and in a more realistic setup. Reference [1] Ashbrook D, Starner S (2002) Learning significant locations and predicting user movement with GPS. In: Proceedings of the 6th IEEE International Symposium on Wearable Computers, Seattle, WA, 7 10 October 2002 [2] Chung, J. (2006) Will You Help Me - Enhancing personal safety and security utilizing mobile phones. Master Thesis, MIT Media Laboratory, [3] Golledge, R. G. (Ed.). (1999) Wayfinding behavior: Cognitive mapping and other spatial processes. Baltimore: Johns Hopkins. [4] Kang, J. H., Welbourne, W., Stewart, B., Borriello, G., (2004) Extracting places from traces of locations, Proceedings of the 2nd ACM international workshop on Wireless mobile applications and services on WLAN hotspots, October 01-01, 2004, Philadelphia, PA, USA

11 11 [5] Krumm, J. and Horvitz, E., (2006) Predestination: Inferring destinations from partial trajectories. In Ubicomp 2006, pages , [6] Liao, L., Fox, D., and Kautz, H., (2004) Learning and inferring transportation routines. In Proc of the 19th Natl Conf on AI, [7] Liao, L.; Fox, D.; and Kautz, H. (2005) Locationbased activity recognition using relational markov networks. In Proceedings of the International Joint Conference on Artifical Intelligence (IJCAI). [8] Lynch, K., (1960) The Image of the City. Cambridge, Massachusetts: The MIT Press. [9] Marmasse, N., Schmandt, C., Location-aware information delivery with commotion. In Proc. HUC 2000, Bristol UK, (2000). [10] Marmasse, N., Schmandt, C., (2002) A User- Centered Location Model Personal and Ubiquitous Computing 2002, p [11] Marmasse, N. (2004) Providing Lightweight Telepresence in Mobile Communication to Enhance Collaborative Living. Ph.D. dissertation, MIT Media Laboratory, [12] Patel, K., Chen, M., Smith, I., Landay, J., (2006) Personalizing Routes, In Proc. UIST 06, 7-5