Safe Walking Route Recommender Based on the Fall Risk Calculation Using a Digital Human Model on a 3D Map

The recent increase in sedentary lifestyles has highlighted the importance of physical activity in our daily lives. Although fall accidents are a common cause of injuries among the elderly, walking is one of the simplest physical activities available. Therefore, walking route selection and identification of walking behavior in different scenarios remains a widely researched area. This paper is about developing a walking-route recommender system that proposes a safe route to the user and reduces the probability of fall accidents. We determine route candidates after the start and endpoints are fed to the system. Then the tripping risk involved in each route candidate is identified by performing several walking simulations of a digital human model on a three-dimensional laser-scanned point cloud. Finally, a safe route is determined after considering the user’s age and the tripping risk. It was confirmed that tripping risk could be competently estimated according to the considered age group.


I. INTRODUCTION
Since "stay home" is recommended worldwide to prevent COVID-19 infection, the physical movement has drastically reduced. Health-care workers and medical experts recommend walking as a critical determinant for alleviating the onset of lifestyle-related diseases. Knowing the value of mobility for a better Quality of Life (QoL) can increase motivation to perform at-home exercises [1]. In Japan, walking has been embraced by individuals of all age groups but is especially popular among those aged 50 years and the elderly [2]. Walking is a routine exercise, but it is challenging to maintain the motivation for such a time-consuming and monotonous activity. The motivation for walking can be maintained by (a) ensuring an appropriate amount of exercise, (b) alerting users of the beneficial effects of training, and (c) providing a safe route. In a related research (a), Claes et al. developed a smartphone platform to present the degree of physical burden during exercise to users [3]. This degree of burden was estimated from acceleration and heart rate data measured through wearable devices attached to the chest. In another research to promote awareness of training effects (b), Hargreaves et al. performed a walking program for 12 weeks. They presented the target number of steps set according to the individual based on the individual's record of past steps [4]. An increase in self-efficacy was observed as the average number of steps per day increased. In another research to support safety assurance (c), Lee et al. developed a system to predict falls using a smartphone's acceleration sensor [5]. The importance of ensuring safety (c) is exceptionally high because there are several instances where walking is discontinued due to difficulties after a fall.
Fall risk is affected by the walking environment that involves geographical features and human features, such as exercise capacity. In a research on determining the fall risk in an environment, Curl et al. developed a checklist for assessing the road surface for walkability. The checklist items were decided by interviewing older adults in detail where they needed to pay attention while walking and where falling probability was high [6]. In another research on determining fall risk affected by exercise capacity, Begg et al. developed a system to display the minimum toe clearance information to the user by measuring the foot's trajectory while walking in real-time [7]. These researches presented the fall risk to pedestrians to ensure a safe road and to warn of danger. In this situation, although pedestrians can recognize dangerous locations through the warning system while taking a walk, they are not aware of a safe path where they can walk with a lesser risk of falling.
Currently, no work has been done that recommends a safe route to pedestrians before they begin the activity. Therefore, this paper presents a walking-route recommender system that proposes a safe route with lowered fall risk. We performed the route recommendation based on fall risk calculation. The environment map data of the concerned area is used to determine the recommended route. The fall risk is calculated through an existing tripping risk quantification method using a digital human model on a three-dimensional (3D) laserscanned point cloud of the walking environment. The remainder of this paper is organized as follows. The Introduction section is followed by Related work, in which a systematic yet brief literature survey is provided. The following section describes the proposed walking route recommendation method based on fall risk calculation. This section is followed by a use case of the proposed method, and the paper is finally concluded.

II. RELATED WORK
In this section, we report recent research advances related to this paper. We divide the literature into four areas: route recommendation systems, fall risk calculation, and Estimation, using 3D maps for navigation and behavior study of pedestrians.

A. ROUTE RECOMMENDATION
Simulation and recommendation of evacuation methods is a type of route recommendation where public spaces are considered [8]. Such types of researches help design evacuation strategies to reduce casualties and damages in the events of natural or man-made disasters. Other researches have highlighted the importance of walking route recommendations for promoting good health [9]. Recent research has proposed an application-based safe route planning solution using COVID-19 occurrence data for cyclists and pedestrians. This study aims at keeping the users away from virus transmission spots [10]. Environmental factors such as air cleanliness [11], land use mix, infrastructure, aesthetics and accessibility [12], transport carbon footprint [13], and postrain accumulated water [14] and traffic factors [12] such as traffic volume and speed, quality of path surface, and safety level of road crossings have also been made primary factors for route recommendation algorithm design. The Land Use Regression (LUR) model was used to model the concentration of black carbon to be calculated in the route candidates for home-to-school commuting [11]. Other researches have aimed at designing route recommendation algorithms for the specific benefit of Chronic Obstructive Pulmonary Disease (COPD) patients [15], visually impaired users [16], and Patients with Motor Disabilities (PWMD) [17]. Such researches have given more importance to the customization of the algorithm using user-specific necessities. Gharebaghi et al. have proposed a Fuzzy Approach for PWMD user (on manual wheelchair)-specific route recommendation in which the user's confidence while dealing with obstacles such as uneven pavement and slopes is used to personalize the route [17].

B. FALL RISK CALCULATION AND ESTIMATION
Fall risk calculation is another primary aspect considered in this paper. Therefore, we report the recent works that have significantly contributed to this area. Three categories of researches have evolved, namely: reasons contributing to falls, fall prevention strategies, and fall risk factors [18]. Identification of the risk factors that contribute to falls helps us design optimum fall prevention strategies. Jensen et al. have categorized these risk factors as extrinsic and intrinsic (health-related) [19]. Extrinsic risk factors include the environmental conditions present in the individual's surroundings, such as clutter, land or floor elevation patterns, loose rugs leading to tripping hazards, lack of railings or grab bars to hold in stairs and bathing areas, and poor lighting. Intrinsic factors include the individual's mental and physical condition. Assessment models such as Home Assessment Profile (HAP) [20], cumulative risk score [21] and FROP-Com (Fall Risk for Older Adults in the Community) [22] were developed for risk calculation by considering the extrinsic factors. A fall risk assessment algorithm for the Japanese inpatient population was proposed recently by considering nine individual intrinsic factors that contribute to increased risk of falls [23]. Fall risk assessment can be done in static and dynamic situations. The intrinsic factors are dynamic since the individual's condition changes as they progress in the environment. Simulation techniques, on the other hand, can predict fall risk in 'what-if' situations [24]. Such simulation is beneficial in evacuation scenarios where environmental factors play a significant role.

C. USING 3D MAPS FOR NAVIGATION
3D digital maps are becoming increasingly on-demand for applications involving smart cities, high-tech living, and autonomous navigation. These are the next step of digital cities, and an important tool for storing spatial information [25]. 3D maps have been proved to be more beneficial to identify certain intricate locations for pedestrian navigation when compared to 2D maps. Three main supportive aspects that can be delivered better through 3D maps compared to 2D are self-orientation, navigation decision, and the knowledge of spatial information [26]. 3D maps are generally constructed using camera-based approaches. Novel data management and processing methods are also being introduced. Jeong et al. proposed odometry estimation for live 3D point cloud registration and error minimization to remove moving vehicles and pedestrians from 3D point clouds using 2D semantic segmentation [27]. These processes were achieved to ensure low computational time and efficient 3D map construction.

D. BEHAVIOUR STUDY
Studying the behavior of pedestrians in specific environments helps us design more efficient navigation systems tailored to the needs of the users. Environmental features are the most prominent factors that affect pedestrian behavior. Studies have shown that people prefer straight, direct paths rather than zig-zag paths, irrespective of the energy expenditure involved [28]. Additionally, visual environmental features can influence the walking speed of pedestrians. The walking speed was reported lower in spaces with a higher degree of natural characteristics such as greenery. In such spaces, people also tend to find city noises more annoying as compared to less natural spaces [29]. The noises we encounter while walking also have a significant on our walking speed. People increase their walking speed as they hear traffic noises, whereas relaxation sounds such as birdsong relax the individual and lower their walking speed [30]. However, it was also observed that active commuting (biking and walking) is often evaded due to psychosocial and personal barriers rather than environmental barriers [31]. The usage of smartphones can have an impact on a pedestrian's priorities. In a Korean study, smartphone walkers reported safety and route quality as essential for route selection. On the other hand, walkers without smartphones considered distance and positive walking experience to be more critical [32]. Using smartphones during walking has an apparent effect on the walking speed accompanied by distracting behavior [33]. Accessibility and street connectivity were also strongly correlated to walkability in the pedestrian street by Jabbari et al.. Here, accessibility was calculated by Angular Segment Analysis by Metric Distance (ASAMeD), and street connectivity was calculated through the space syntax measure of connectivity [34]. An agent-based study showed that landmarks tend to improve the correspondence of the user with the navigation routes. Therefore, landmarks play an essential role in route planning and also tend to lead agents towards green areas more frequently [35].

III. WALKING ROUTE RECOMMENDER SYSTEM BASED ON FALL RISK A. SYSTEM CONFIGURATION
This study proposes a system that recommends safe walking routes. At first, the user inserts the start and endpoints of their route. The system determines multiple routes between the two points. For each route, the travel distance and the fall risk by tripping are estimated. Finally, an appropriate and safe walking route is determined for the user based on the fall risk, distance, and user's age. The path structure is acquired from Google Maps, and the roughness present on the surface of the road (which largely determines the fall risk) is acquired from a 3D laser-scanned point cloud of the environment (called 3D map). Figure 1 shows the walking-route recommender system. The system consists of a tripping risk calculator (A1) and a recommender (A2). Part (A1) creates the database of the tripping risk by walking simulation on the 3D map [36]. Part (A2) creates a route candidate group from the start to endpoints based on Google Maps and determines the tripping risk along every route candidate using the tripping database. Finally, the route L a that minimizes the user's tripping risk in their age group is recommended.

B. CALCULATING FALL RISK USING 3D MAPS AND DIGITAL HUMAN MODEL
The tripping risk calculation part (A1) prepares a tripping risk database along each walking route. This database is vital to determine the optimum route that involves the lowest probability of tripping. An existing technique calculates the tripping risk; namely, simulation of the walking situation on a 3D laser-scanned point cloud [36]. Our route recommender system is visualized by a 3D laser-scanned point cloud of the Nihon University campus of Japan. The fall risk calculation method is reported below in further detail.
1) Estimation of the walking surface and path structure from the laser-scanned point cloud This method first includes the extraction of a point cloud representing the walking surface [36]. This extraction is done by the popular region-growing method [37]. Based on the obtained point cloud distribution, the path graph illustrating the environment's route connectivity is constructed. Figure 2 shows the image of an original point cloud and a path graph showing the various route connections of the same point cloud. From this path graph, the walking trajectory from the start point to any target point and the possible walking motions on the 3D point cloud can be simulated [38].

2) Simulation with the Digital Human Model (DHM)
Walking simulation using the DHM is performed on the actual environment model mentioned in the section III-B1 [39]. The DHM had 41 degrees of freedom, and the dimensions of each segment of the body are based on the gait database [40].  The simulation is realized in three steps as proposed in [36]. The first step is to select nodes and edges between q 0 and q e or the start and endpoints, respectively. The second step is to determine the walking trajectory s k based on an optimization algorithm. The third step simulates the walking motion, i.e., full-body motion during the walking trajectory s k . The simulation generates the motions of each step referring to the flat-walking motion data captured in the gait database [40] and the walking-related parameters (step length and toe clearance) [41].

3) Calculation of the tripping rate
The tripping rate is calculated in a Monte Carlo simulation [42], where the walking simulation is executed many times with different walking parameters [36]. In each trial, occurrences of tripping or non-tripping are estimated by simulating the collision between the DHM foot and the walking surface. The toe clearance C t is the difference between the current toe height of the DHM and the height of the road surface where the next step will be placed during the walking simulation. Figure 3 shows the foot of the DHM and the road surface for determining the toe clearance C t in two different situations of step-up and step-down. When the foot of the DHM interferes with the road surface as shown in Figure 3(b), i.e., the condition of C t ≤ 0 is encountered, it is judged as tripping.
The number of tripping occurrences C trip in each trial of walking simulation is calculated by Equation 1. The ratio of the number of tripping occurrences C trip to the total number of simulations N sim is termed as the tripping rate rq 0 , q e ) Figure 3: Conditions for determining falling in simulation while walking from point q 0 to q e . Therefore, the tripping rate r(q 0 , q e ) is given by Equation 2.
The algorithm described in this section is used to calculate the tripping rate involved in a particular walking trajectory between the start position q 0 to the end position q e to quantify the fall risk.

C. RECOMMENDATION OF THE OPTIMUM WALKING ROUTE CONSIDERING TRIPPING RISK
Part (A2) of the proposed system refers to the tripping risk database obtained in part (A1) and recommends route candidates to the user, avoiding routes with high tripping risk as shown in Figure 1. A group of route candidates L is generated from the map data (Google Maps). A route candidate group is termed as L = {L i | 0 ≤ i ≤ N } where L i is the i th route candidate. The total distance to be walked if the i th route candidate L i ∈ L is selected, is represented as d(L i ). The structure of the route candidate group L i is shown in Figure 4. The i th route candidate consists of relay points p j between the start point p 0 and the end point p M such that p j | 0 ≤ j ≤ M . Next, the fall risk t(L i ) along each route L i is calculated based on the tripping risk data obtained during the step (A1). Finally, a safe route with low tripping risk is recommended as the route L a that minimizes the tripping risk. The route recommendation method is detailed below.

1) Generation of the route candidate group L
A route candidate group is prepared to compare the fall risk involved in each route. After the user specifies the start and endpoints (p 0 and p M , respectively) of the walking route, the shortest-route L 0 between p 0 and p M is determined by the shortest-route search method [43]. Moreover, the N th shortest path search method [44] generates route candidate groups L = {L i } in the order of decreasing distance d(L i ).
As shown in Figure 5, all relay points of the shortest route are considered as branch point candidates in the N th shortest path search method [44]. The shortest route search is executed from the start to each branch point and from the branch to the endpoints. The distance covered in each shortest route for all branch points is calculated. The shortest path among these moving distances is finally selected. This process is repeated N times to find the N th shortest path. The coordinate system represents the relay points of the route candidates L i and the walking simulation trajectory s k in the tripping risk database. The relay points p j along the route candidates L i (L i ∈ L) are expressed by their longitude and latitude in the geographical coordinate system. Alternatively, the start point q 0 and endpoint q e along the trajectory s k created in section III-B2 are represented in the Euclidean coordinate system of the point cloud data. Thus, these data are converted to a common rectangular coordinate plane (9 series) [45].
(a) Transformation from geographical coordinates The geographical coordinate system is converted into a rectangular coordinate plane through a conventional coordinate transformation formula [45]. This helps to express the relay pointsĝ j along route L i in rectangular coordinates.
(b) Transformation of point cloud coordinates The next step is to construct the coordinate transformation matrix T . Eight landmark points (such as corners) are selected from the 3D map and designated as x i (i = 0, 1, 2, · · · , 7). The corresponding eight points y i (i = 0, 1, 2, · · · , 7) on the rectangular plane are also manually selected from Google Maps. Now the X and Y coordinates are designated as X = [x 0 , x 1 , x 2 , · · · , x 7 ], Y = [y 0 , y 1 , y 2 , · · · , y 7 ]. A coordinate transformation matrix T is calculated from x i to y i to satisfy y i = T x i . Using the obtained matrix T , the coordinates of the startq 0 and the end pointsq e of the walking trajectory s k are expressed in the rectangular plane coordinate system as follows: This coordinate transformation gives the route candidate L = {L i } (L i = {p j }), and the walking simulation trajectory group S = {s k | 0 < k < N s } in the rectangular plane coordinate system is also obtained. Each walking simulation trajectory can be represented as s k = {q l | 0 ≤ l ≤ e}.
3) Determining the walking trajectory s k belonging to the route L i The walking trajectories s k belonging to each route candidate L i are searched from the walking trajectory group S = {s k } of the tripping risk database. As shown in Figure 6, the start and end points of the walking trajectory s k are set toq 0 and q e , respectively, and the relay points of the routes are set tô p j . After ensuring thatq,p j , andp j+1 satisfy Equations 4 and 5, the distance d between the line segment connectinĝ p j ,p j+1 and the pointq 0 orq e is calculated by Equation 6. In Equation 6, n is the normal vector of the line segment connectingp j ,p j+1 and the point on the walking trajectory s k (q 0 orq e ).
As shown in Figure 7, Equations 4 and 5 are satisfied when the foot of the perpendicular line from the pointq to the line segment connectingp j andp j+1 coincide with the line segment itself. When the distance d from route L i is less than or equal to the threshold l for pointsq 0 andq e , then trajectory s k belongs to L i as shown in Equation 7.
The walking trajectory s k belonging to the route L i is defined as a walking trajectory group s L i = {s k }.  The tripping rate r(q 0 ,q e ) is obtained from the tripping risk database created in section III-B by specifying the walking trajectories s L i on the route L i and the user's age (younger or elderly people). The largest value among the tripping rates {r(q 0 ,q e )|s k ∈ L i }(q 0 ,q e ∈ s k ) of the walking trajectory group L i = {s k } is considered as the tripping risk t(L i ) of the particular route L i . The fall risk is averaged for each age group of users (younger or elderly people) to determine the tripping rate corresponding to the age group.

5) Determining the recommended route La
Finally, the recommended walking route L a that minimizes the fall risk among the route candidate groups L i is determined, where t is the calculated fall risk in the route candidate group L i as mentioned in section III-C4.
L a = arg min

IV. CASE STUDY FOR VALIDATING THE PROPOSED METHOD A. CONDITION
The walking simulation for calculating the fall risk was performed using the DHM and a 3D laser-scanned point cloud of the environment. The conditions and results are described in this section.

1) Selection of the start and end positions
The first step is to input the start p 0 and end p M positions of the desired walking route and the user's age. The two pairs of start and end positions selected for this case study are shown in Table 1 and Figure 8. The optimum route L a was recommended for both young and older adults in each condition. An extensive route search follows this. 2) Point cloud data for calculating the fall risk A 3D laser-scanned point cloud published by Nihon University was used for calculating the tripping probability. This  The walking simulation for calculating the fall risk was performed using a DHM [40] under the conditions shown in Table 2. The motion capture data of seven young and seven older adults were used for the simulation. For each subject, the walking simulation was repeated 500 times in the Monte Carlo simulation. The tripping rate was calculated with N = 3500 (7subjects * 500) simulations for each young and elderly group. The parameter settings of sub. 1-7 correspond to the DHM of a young person, whereas those in sub. 8-14 correspond to the DHM of an elderly individual.
To represent the route recommendation mathematically, considering 57 pairs of manually specified start q 0 and end points q e , the system generates the walking trajectory group S = {s(q 0,i , mq e,i )}(i ∈ [1,57]) to simulate the walking motion by DHM on each trajectory. The tripping rate is calculated for the young (sub. 1-7 in Table 2) and the elderly (sub. 8-14 in Table 3) people separately. Given the walking trajectory s(q 0,i , q e,i ) and the age group A, we could obtain the tripping rate r A (q 0,i , q e,i ) on the specific walking trajectory.

B. RESULT
This section shows the results of the proposed walking route recommendation system. Figure 10 shows a walking simulation using the reported DHM. The calculation results of the routes and their fall risk are shown in Figures 11 and 12. The points with the highest fall risk are marked by a red circle in Figures 11 and 12. Tables 3 and 4 represent the distance d(L i ) and the estimated fall risk r for each route L i and age group. Hence, the safest route for young and older adults can be reported as L a = L 1 in condition 1 and L a = L 0 in condition 2.   Tables 3 and 4, the recommended route can minimize the fall risk r of both young and elderly people were L a = L 1 in condition 1, and L a = L 0 in condition 2. It has been reported in previous studies that routes that include steps of height 50 mm have an average fall risk of 23 % for young people and 30% for elderly [36]. The fall risks of routes L 1 and L 2 exceeded 23 % for people of both age groups in condition 2, and both routes included occurrences of step height exceeding 50 mm. Hence, it can be said that our proposed fall risk calculation method is reliable. Comparing the fall risk by age group in the same route L i , the fall risk during walking increases with age as shown in Tables 3 and 4. In Japan, people aged fifty and above are approximately seven times more likely to trip than people in their twenties [46]. This trend was mirrored in the calculated tripping risks r.  The recommended routes for conditions 1 and 2 were L 1 and L 0 , respectively. In condition 1, the fall risk of L 1 was 2.2 % for young and 4.4 % for elderly subjects. The route with the lowest fall risk was recommended because routes with very high and low fall risks were included in the route candidates. However, suppose the user gives preference to long-distance coverage for their walk. In that case, the recommender system can select moderate fall risk and more extended distance coverage, which makes L 2 as the optimum selection for condition 1. For condition 2, the fall risk of L 0 was 12.8 % for young and 24.2 % for elderly subjects. The route with the lowest fall risk was selected. Unlike condition 1, in the case of condition 2, all other route candidates show high fall risk, especially for the elderly. Therefore, the best route recommendation is L 0 , no matter the distance preferences.

As shown in
In condition 1, if we observe the locations projecting the most significant fall risks (marked with a red circle in Figure 11), routes L 0 and L 2 included both curbs and road bumps or unevenness (see Figure 13), and ditches immediately before the curb. The fall risk was higher at these sites than along L 1 because the step height was relatively large. Therefore, it was confirmed that routes L 0 and L 2 with an estimated high fall risk were avoided, while the safest route L 1 was proposed to the user. Figure 14 shows the difference in fall risk between the young and the elderly from Tables 3 and 4. In condition 2, the difference in fall risk was 11.4 % for L 0 , 22.9 % for L 1 , and 10.1 % for L 2 . Although the recommended routes for the young and elderly subjects were the same in the case study results reported in this paper, routes with significantly different fall risks for the young and the elderly subjects were confirmed. In route L 1 , older adults' fall risk was significantly higher than that of young people compared with other routes. Although L 1 can be recommended for young people because the fall risk is low, it should not be recommended for older adults because of high fall risk, thus, an unsafe route. The efficiency of route selection can be further improved by adopting a route such as L 1 , as the youth route candidate. In the case of older adults, excluding such routes from the candidate routes group would prevent dangerous route recommendations. A safer and more age-appropriate route can be proposed by choosing the route based on the fall risk linked to each age group and the distance covered.
Furthermore, suppose an individual (preferably young) is looking for a challenging activity, apart from fall-risk reduction. In that case, the proposed methodology can escalate motivation to exercise by recommending challenging routes to strengthen the user's body functionality. In the future, we aim at performing route recommendations considering the user's physical ability irrespective of age by evaluating the user's ability through pre-tests and measurements. In addition, the route's safety can be evaluated more realistically by adding indices such as the existence of stairs, road surface conditions, post-rain water accumulation, presence of a pedestrian crowd, adequate brightness, and collision probability with vehicle traffic.

V. CONCLUSION
In this paper, we proposed a walking route recommender system to support pedestrians. We confirmed that the system could propose a safe route by calculating the tripping rate through walking simulation of a DHM on a 3D laserscanned point cloud. The system could also propose an ageappropriate route by calculating the tripping risk according to the age group. PRIYANKA RAMASAMY recieved a B.E degree in Electronics and Communication Engineering from RVS faculty of Engineering, affliated to Anna university, India, in 2014. In 2021, she acquired M.E degree and currently pursuing PhD in System Cybernetics through Taoyaka program under the supervision of Prof. Yuichi Kurita with the Biological Systems laboratory in Hiroshima University. Her research interest includes machinehuman interaction through gamification, physical rehabilitation, human augmentation through soft robotics, learning of physical and psychological effects through wearable assistive technologies.