pSafety: A Collision Prevention System for Pedestrians Using ...

15 downloads 36310 Views 409KB Size Report
pSafety, which instantaneously informs pedestrians and drivers of the potential ... has become non-ignorable with high usage of cell phones. [3]. Although ... is on the phone. WiFi-. Honk [11] modifies Android code to make each pedestrian's.
pSafety: A Collision Prevention System for Pedestrians Using Smartphone Chi-Han Lin† , Yi-Ting Chen† , Jyun-Jie Chen† , Wen-Chan Shih] , and Wen-Tsuen Chen†] † Department of Computer Science, National Tsing Hua University, Hsin-Chu, 300, Taiwan ] Institute of Information Science, Academia Sinica, Nankang, Taipei 115, Taiwan E-mail:[email protected], [email protected], [email protected], [email protected], [email protected] Abstract—We implement a collision prevention system, called pSafety, which instantaneously informs pedestrians and drivers of the potential threatening accidents. Unlike other systems, pSafety alerts pedestrians of threatening vehicles coming from not only the line-of-sight, but also non-line-of-sight due to obstructions of the wall corner or other vehicles. pSafety collects GPS information from smartphones of pedestrians and vehicle drivers through mobile networks. The main challenge of pSafety is that current smartphones demonstrate larger distance errors on the order of a few meters due to its intrinsic low-cost GPS receivers. To address the impact of large-error positioning to pSafety, we regard each participant on the map as a sector that indicates a predicted location. We subsequently design the Sector Overlap Detection Algorithm, called SODA, to detect whether two sectors are overlapping in time O(1). To avoid warning fatigue, we additionally provide a threat ranking method to evaluate the degree of risk for each potential collision event. Through our designed App, pedestrians and drivers both could receive a clear view of potential risks and then take proper actions to avoid accidents. In our implementation, we show that pSafety rapidly informs participants (i.e., pedestrians and drivers) and provides each participant a sufficient response time to avoid collision. Keywords—collision prevention system, sector overlap detection, collision threat ranking.

I.

I NTRODUCTION

Traffic accidents caused by distracted pedestrians have proliferated in recent years. According to statistics, 32% of pedestrians listen to music, 14% text, 9% browse the Internet, 7% use social media, and 3% play games or watch TV/videos while crossing roads [1]. In addition, from 2009 to 2012, pedestrian fatalities due to traffic accidents have been rising at an annual rate of 4.9% [2]. Econometric analysis has also shown that the life-taking impact due to distractions has become non-ignorable with high usage of cell phones [3]. Although Collision Prevention Systems (CPS), Collision Avoidance Systems (CAS), or Pre-Collision Systems (PCS) have been developed by many vehicle manufacturers for several years [4], none of them provides active alerts for distracted pedestrians in order to avoid traffic collisions. Additionally, current CPSs suffer from a main challenge: The moving object must be within the line-of-sight of vehicles. Since the view to moving objects may be covered by obstacles on the road, e.g., moving/parked truck/bus and wall corner, current detection technologies, such as radar, sonar, laser, and camera, cannot operate properly. Therefore, some studies adopt cooperative perception via vehicle-to-infrastructure and vehicle-to-vehicle communications to exchange information with nearby vehicles and roadside units to recognize potential

collision events [5]. However, these solutions usually require extra device support such as Dedicated Short Range Communications (DSRC) [6], which is uncommon for a commodity smartphone or tablet. Sharing position among neighboring users is another solution for pedestrian safety. Nevertheless, most of this type of works also require extra devices to get accurate positioning information, like Volvo cyclic helmet [7], which is not convenient for pedestrians. Although using the intrinsic GPS receiver in smartphone is an intuitive solution, this suffers from another challenge. That is, the GPS receiver of a commodity smart device is usually low-cost, so that the positioning information is inaccurate on the order of meters. Thus, some works focused on how to improve the accuracy of smartphone GPS receiver [8]. However, these works require the root privilege of smartphone, which is unfriendly for general users. In this paper, we implement pSafety, which adopts the intrinsic GPS receiver of smartphone and could instantaneously alert a pedestrian distracted by the smartphone to potential collision events via App, even the pedestrian is within the non-line-of-sight of vehicles. In order to reduce the impact of large-error positioning in smartphones, we regard the location of each user as a sector which indicates a predicted area. We then design a Sector Overlap Detection Algorithm, called SODA, to detect whether two sectors are overlapping in time O(1). To the best of our knowledge, this paper is the first work that systematically investigates the sector overlap problem. We also take warning fatigue into consideration by providing a threat ranking method to evaluate the degree of risk for each potential collision event. Through our experiments, even the relative location errors between two smartphones is about 7m, the positive rate (i.e., the rate that pSafety alerts participants around all experiments) is high when the pedestrian is near danger, as well as, becomes low when the pedestrian is in relatively far. The results also indicate that warning fatigue problem impacts pSafety very limited. The organization of this paper is summarized as follows. Section II reviews state-of-the-art technologies about CPS and other pedestrian safety solutions. The system overview of pSafety, scenario, problem definition, proposed algorithm, SODA, and the threat ranking method are described in Section III. Section IV shows our experimental results, and Section V concludes this work. II.

R ELATED W ORK

In this section, the state-of-the-art CPS and other pedestrian safety solutions are reviewed as follows:

CPS: Most CPSs are designed for vehicles, which adopt multiple detection technologies such as infrared, radar, laser, camera, and so on to identify pedestrians in front of vehicles [4]. Generally, they first calculate the distance to the object and relative velocity via these detection technologies, and then estimate the collision possibility. However, current CPSs are effective only if the vehicle speed is lower than 30 Km/s [9], and require that the object is within the line-of-sight of the vehicles. For non-line-of-sight scenarios, Ko-PER [5] applies vehicle-toinfrastructure and vehicle-to-vehicle communication to inform nearby vehicles of potential danger. Nevertheless, both CPSs and Ko-PER cannot actively inform pedestrians of vehicles with high collision risk because current smartphones do not support Dedicated Short Range Communications (DSRC) [6]. Other solutions for pedestrian safety: WalkSafe [10] also requires the line-of-sight between a pedestrian and vehicle, which adopts smartphone camera to recognize vehicles coming from pedestrian side. However, WalkSafe only works when the pedestrian is crossing the road and is on the phone. WiFiHonk [11] modifies Android code to make each pedestrian’s smartphone send Beacon frame periodically. Each vehicle receiving Beacon then evaluates the collision possibility and informs both driver and pedestrian. For WiFi-Honk, the root privilege of smartphone is required, and Beacon frame of IEEE 802.11 needs to be modified to involve pedestrian location, velocity, and orientation information. LookUp [12] is designed for technology-distracted pedestrians, which provides a different perspective for pedestrian safety. It uses wearable device, namely intelligent shoes, to detect whether a pedestrian is entering a road entry from sidewalk. Subsequently, LookUp reminds the distracted pedestrian to look up right before entering a road entry through smartphone App. Although this work does not focus on collision detection, it could combine with pSafety to further reduce warning fatigue. III.

T HE P ROPOSED C OLLISION P REVENTION S YSTEM

In this section, the proposed collision prevention system, pSafety, is detailed. We first describe the system overview, including requirements and components. Second, we define our problem and list all of the used notations in this paper. Subsequently, the overlap detection problem of any two sectors, called SODA, is comprehensively discussed in terms of four cases. A threat ranking algorithm is then provided when overlap occurs. Finally, map partitioning is adopted to reduce the total execution time of SODA. A. System Overview Fig. 1 shows the components of proposed system including smartphones and local servers. We assume that vehicle drivers and pedestrians carry their smartphones and activate the GPS. The participating smartphones upload GPS locations periodically to nearby local servers. A local server can be installed in a government institute, such as a police office or a local traffic control center. When a GPS location of a participating smartphone is updated, the local server executes SODA and the threat ranking method to update the degree of collision risk. Finally, local servers notify all participating smartphones and alarm participating drivers and pedestrians with high collision risk via smartphone App. The designed smartphone App performs a navigation mode and a background mode

Fig. 1: System overview

for vehicle drivers and distracted pedestrians, respectively. Through this App, the direction of the potential threat and the degree of risk could be shown even if distracted pedestrians are using other Apps. B. Problem and Notation Definition Here, we show how to perform user location prediction in a form of sector and define all used notations. A 2-D plane in outdoor environment is taken into consideration, where each GPS location is converted to Cartesian coordinate and is represented as a point on the plane. Let si,t = (xi,t , yi,t ) denote the GPS location of user i on the map at time t, where xi,t and yi,t are Cartesian coordinate of x-axis and yaxis, respectively. A local server then uses (1) to calculate the instantaneous velocity vector of user i at time t, denoted by ~vi,t . ~vi,t =

(xi,t − xi,t−1 , yi,t − yi,t−1 ) si,t − si,t−1 = . t − (t − 1) t − (t − 1)

(1)

Note that, in (1), we only introduce a simple velocity estimation, which may be improved by applying other estimation methods such as Kalman filter. Let T denote the prediction time interval, and we then have the total distance ri,t between si,t and si,t+T , where user i may locate within T , as formulated in (2). ri,t = T · |~vi,t |. (2) In our implementation, due to the consideration of human response time, T is empirically set to 5s. Since a user may shift from its original direction within T , the angle θi,t that user i shifts from its original direction within T is additionally estimated by (3). α θi,t = arctan( ) (3) |~vi,t | where α is an empirical system parameter equal to 2.334 according to our experiments. Consequently, a predicted sector of user i at time t could be constructed, denoted by Ai,t , whose centre is at si,t , central direction is toward ~vi,t , radius equals ri,t , and sector angle equals 2θi,t . Fig. 2a shows an example of a predicted sector. If a sector Ai,t overlaps another sector Aj,t , user i may collide with user j within T . The next shows how to detect overlap between any two sectors by the method of exhaustion.

T vj,t

T vi,t

T vj,t

sj,t rj,t

sj,t

Ai,t ri,t

θi,t

ri,t

θi,t θi,t

ri,t θi,t

si,t

c1

T vi,t

θiaa

T vi,t

si,t

si,t

user i

(a) Case 1: inside

θjbb θj,t

ri,t

c2

ri,t θi,t sj,t

sj,t

(b) Case 2: edge intersection

si,t

(c) Case 3: arc intersects with edge

(d) Case 4: arc intersection

Fig. 2: The four cases of overlapping sectors

C. Sector Overlap Detection Algorithm (SODA) We separately investigate four possible cases that two sectors overlap. Note that SODA is executed only if |~si,j,t | ≤ ri,t + rj,t , where ~si,j,t = sj,t − si,t = (xj,t − xi,t , yj,t − yi,t ) = −~sj,i,t denotes the relative location vector from user i to user j at time t. Otherwise, sector Ai,t and Aj,t do not overlap. The comprehensive procedure of SODA is shown in Algorithm 1. Case 1: Inside- This case indicates that a user location, say sj,t , is located inside Ai,t . Fig. 2a shows an example of this case. Two conditions need to be checked: 1) Whether the distance between si,t and sj,t is longer than ri,t ; 2) Whether the angle between ~vi,t and sj,t − si,t , denoted by θ0 , satisfies with (4). θi,t

~vi,t · ~si,j,t ). ≥ θ (~vi,t , ~si,j,t ) = arccos( |~vi,t ||~si,j,t | 0

(4)

Case 2: Edge intersection- As shown in Fig. 2b, this case indicates that at least one edge of Ai,t intersects with at least one edge of Aj,t . Given si,t = (xi,t , yi,t ) and ~vi,t , we have two line equations of two edges on Ai,t , denoted by li,t,1 and li,t,2 , through (5) and (6).

2 y = yi,t + aj ri,t /(1 + a2j ),

(8)

aj = cos(arccos(~vj,t ) ± θj,t ).

(9)

where ae

Let C denote the set of all cross-points in this case. Conseae quently, this case occurs if θi,t ≥ θ0 (~vi,t , (sae k − si,t )) and sk ae is on the edge of Aj,t , ∃sae ∈ C . k Case 4: Arc intersection- This case indicates that the arc of Ai,t intersects with the arc of Aj,t , as shown in Fig. 2d. Note that this case occurs only if |ri,t − rj,t | ≤ |~si,j,t | ≤ ri,t + rj,t . We first use (10) to calculate θiaa according to the law of cosines since the length of each edge of 4si,t c1 sj,t is known. θiaa = arccos(

2 2 |~si,j,t |2 + ri,t − rj,t ). 2ri,t |~si,j,t |

(10)

Then, the angle range of circle intersection relative to ~si,j,t is between ±θiaa . Finally, if θ0 (~vi,t , ~si,j,t ) − θiaa − θi,t < 0, this case occurs. Since each loop in Algorithm 1 is performed up to 4 times, i.e., line 8 and 16, each case in the sector overlap problem takes O(1) time. Therefore, the total running time of Algorithm 1 is O(1).

li,t,1 : y − yi,t = cos(arccos(~vi,t ) + θi,t ) · (x − xi,t ).

(5)

D. Threat ranking method

li,t,2 : y − yi,t = cos(arccos(~vi,t ) − θi,t ) · (x − xi,t ).

(6)

For reducing warning fatigue, ranking the degree of collision risks is necessary to filter out some alarms with low collision risk. Thus, we propose the threat ranking method that assigns a degree of collision risk F to each overlapping event detected by SODA, where F ∈ R, 0 ≤ F ≤ 1 and is defined in (11).

Similarly, we have lj,t,1 and lj,t,2 and could calculate all crosspoints. Note that there are at least two cross-points even if parallel lines exist. For this case, let C ee and s0k denote the set of all cross-points and the kth cross-point among all crosspoints, respectively. This case occurs if |s0k − si,t | ≤ ri,t , |s0k − sj,t | ≤ rj,t , θi,t ≥ θ0 (~vi,t , (s0k − si,t )), and θj,t ≥ θ0 (~vj,t , (s0k − sj,t )) hold simultaneously, ∃s0k ∈ C ee . Case 3: Arc intersects edge- This case indicates that an arc of a sector intersects with an edge of another sector. Fig. 2c shows that the arc of Ai,t intersects with one edge of Aj,t . We first check whether the distance between si,t and line lj,t,m is less then ri,t , where m ∈ {1, 2}. Subsequently, the circle equation of Ai,t ’s arc and line equation of lj,t,m are used to calculate cross-points. A cross-point location in this case, denoted by sae k = (x, y), could be calculated by (7) and (8). 2 x = xi,t + ri,t /(1 + a2j ),

(7)

F = 1 − exp(−

max(vti,j , 0) · ∆di,j t ), |~si,j,t |

(11)

In (11), vti,j denotes the relative velocity between user i and j at time t. vti,j > 0 indicates that user i is approaching user j, whereas vti,j < 0 indicates leaving. Here, we only take the case of mutual approaching into consideration and therefore adopt max(vti,j , 0). ∆di,j t denotes the estimated longest duration that user i or j locates in the overlapping area of two sectors, which are respectively defined in (12) and (13). ~vi,t · ~si,j,t ~vj,t · ~sj,i,t vti,j = + . (12) |~si,j,t | |~sj,i,t |

Algorithm 1 Sector Overlap Detection Algorithm (SODA) Input: Sector Ai,t including si,t , ~vi,t , ri,t , and θi,t Sector Aj,t including sj,t , ~vj,t , rj,t , and θj,t Output: true or false 1: if |si,t − sj,t | > ri,t + rj,t then 2: return false 3: end if 4: if (|~ si,j,t | ≤ ri,t and θi,t ≥ θ0 (~vi,t , (~si,j,t ))) or (|~si,j,t | ≤ rj,t and θj,t ≥ θ0 (~vj,t , ~sj,i,t )) then 5: return true //Case 1 6: end if 7: Calculate li,t,1 , li,t,2 , lj,t,1 , lj,t,2 by (5) and (6) 8: for all m ∈ {1, 2}, n ∈ {1, 2} do 9: if li,t,m ∦ lj,t,n then 10: s0 ← the cross-point of li,t,m and lj,t,n 11: if |s0 − si,t | ≤ ri,t and |s0 − sj,t | ≤ rj,t and θi,t ≥ θ0 (~vi,t , (s0 −si,t )) and θj,t ≥ θ0 (~vj,t , (s0 −sj,t )) then 12: return true //Case 2 13: end if 14: end if 15: end for 16: for all p ∈ {i, j}, m ∈ {1, 2} do 17: q←i 18: if p = i then 19: q←j 20: end if 21: if distance(sp,t , lq,t,m ) ≤ rp,t then ae 22: sae p,1 , sp,2 ← cross-points of lq,t,m and the circle equation of Ap,t by (7), (8), and (9) ae 23: if θp,t ≥ θ0 (~vp,t , (sae p − sp,t )) and sp is on the edge ae ae of Aj,t , ∃sae ∈ {s , s } then p p,1 p,2 24: return true //Case 3 25: end if 26: end if 27: end for 28: if |ri,t − rj,t | ≤ |~ si,j,t | ≤ ri,t + rj,t then 29: Calculate θiaa and θjaa according to (10) 30: if θ0 (~vi,t , ~si,j,t )−θiaa −θi,t < 0 or θ0 (~vj,t , ~sj,i,t )−θjaa − θj,t < 0 then 31: return true //Case 4 32: end if 33: end if 34: return false

∆di,j t = max(

T vti,j − |~si,j,t | , 0). min(|~vi,t |, |~vj,t |)

(13)

The design concept of (11) is inspired by Poisson process. Since |~si,j,t |/ max(vti,j , 0) represents the total moving time spent from the current location to the collision point according to the current relative velocity and distance, the item max(vti,j , 0)/|~si,j,t | could be considered as the average collision times per unit time. Thus, the degree of collision risk F could be imagined as the probability of at least one collision in this overlapping event. Finally, a threshold Fthr is empirically set to 0.3 in our implementation so that a local server reports an overlapping event whose F is larger than Fthr .

vehicle 3

vehicle 1 G D

H

A

pedestrian 1 E

B

vehicle 2 vehicle 4

2Tvmax I

Vehicle

Pedestrian

F

C

2Tvmax

Fig. 3: An example of map partitioning

E. Map partitioning In order to simultaneously serve numerous vehicles and pedestrians, inspired by [13], map partitioning is adopted in pSafety to reduce the total execution times of SODA. Since each vehicle-pedestrian pair should be checked by SODA, when one of them updates its location, SODA has to be executed O(Nv ) or O(Np ) times, where Nv and Np are respectively the number of vehicles and pedestrians stored in a local server. This inefficiency could be improved by partitioning the map into several square regions with equal size. When a user updates its location, the local server only checks within its neighboring regions. For example, in Fig. 3, if pedestrian 1 updates its location, only vehicles in region A-I, i.e., vehicle 1 and 2, are checked by SODA. However, the width of a square region should be properly configured to avoid that two sectors overlap like vehicle 3 and 4 in Fig. 3. Therefore, the width of a square region should be set to 2T vmax , where vmax is the maximum speed of objects, to guarantee comprehensive detection. In our implementation, vmax is set to 50m/s, and thus the width of a square region equals 500m. IV.

I MPLEMENTATION AND P ERFORMANCE E VALUATION

In this section, we evaluate pSafety’s performance through empirical experiments on the road. We adopt HTC One M8 and SONY Xperia Z3 compact to be the smartphones on a vehicle and a distracted pedestrian, respectively. The positioning error between the two smartphones is 7-8m. The local server and smartphone App are separately implemented by JAVA and Android. Since crossing the road is the most dangerous situation for pedestrians [14], the scenario of crossing the road is taken into consideration, as shown in Fig. 4. The distracted pedestrian is attempting to cross the road and standing in front of a parked vehicle, which is the non-line-of-sight for the moving vehicle. The average speed of the moving vehicle is 55km/h, and the initial distance between the collision point and the moving vehicle is 200m. Each experiment round is ended when the moving vehicle moves over the predicted collision point. The experiment results are collected from 13 experiment rounds. Fig. 5 shows pSafety’s response time and the degree of collision risk F defined in (11). The response time is defined by the time interval that the distracted pedestrian could be alarmed before the collision occurs. In this measurement, the

distracted pedestrian

parked vehicles

V.

sidewalk predicted collision point

moving vehicle road

Fig. 4: The experimental scenario: crossing the road

The degree of collision risk, F

1 0.9 0.8 0.7 0.6

exp1 exp3 exp5 exp7 exp9 exp11 exp13

0.5 0.4 0.3 0.2 0.1

exp2 exp4 exp6 exp8 exp10 exp12

C ONCLUSIONS

We propose pSafety, a GPS-based CPS for distracted pedestrians, to inform distracted pedestrians and vehicles of potential collision events in time via our smartphone App. We first construct a sector that represents a predicted location for each vehicle and pedestrian through collecting GPS from smartphones. The sector overlap problem is then comprehensively investigated, and SODA is thus proposed to detect whether two sectors overlap on the map, which runs in O(1) time. To reduce warning fatigue, We also propose the threat ranking method to evaluate the degree of risk of each overlapping event. Benefiting from map partitioning, pSafety could serve numerous vehicles and pedestrians more efficiently and instantaneously alerts distracted pedestrians via smartphone App. Through our App, the relative direction of vehicles with high risk will be shown on the screen. In our experiments, pSafety achieves high positive rate and rapidly alerts at 3-4s before collision happens.

0 0

1

2

3

4

5

Time to collision (s)

R EFERENCES

Fig. 5: The system response time and the degree of collision risk F

[1]

[2] [3]

distance between the distracted pedestrian and the predicted collision point is 1m, which pSafety should alarm the distracted pedestrian due to the short distance. Let T C denote the time point that the vehicle arrives the predicted collision point. R denote the time point that the nth event report of the Let Tm,n mth experiment is received by the pedestrian’s smartphone, where m, n ∈ N. Each data point in Fig. 5 then represents the R time interval between T C and Tm,n . Consequently, considering the time of human response and taking proper actions, Fthr is set to 0.3 since participants could be alerted about 3-4s before the collision occurs among all experiments. Therefore, pSafety’s response time could achieve about 3-4s before the collision occurs. We then demonstrate pSafety’s detection rate, as shown in Table I. In this measurement, we change the distance between the pedestrian and the predicted collision point from 1m to 5m. Due to the inaccurate positioning of smartphone GPS, the distracted pedestrian may be alarmed even the distance is long enough, e.g., 4m and 5m. However, the positive rate of short distance cases, i.e., 1m and 2m, achieves high detection rate. Since this rate rapidly decreases when the distance increases, we could conclude that warning fatigue is not a tremendous problem of pSafety.

[4]

[5]

[6] [7] [8]

[9]

[10]

[11]

[12]

TABLE I: The positive rate under different distance between the pedestrian and the collision point distance 1m 2m 3m 4m 5m

positive rate 100% 92.31% 76.92% 53.85% 46.15%

max. F 0.992 0.920 0.879 0.672 0.837

median F 0.721 0.703 0.621 0.566 0.529

min. F 0.306 0.371 0.364 0.359 0.322

[13]

[14]

“Why did the (young) pedestrian cross the road? to talk on the phone, text and watch videos, ford survey shows,” 2015, available online at http://goo.gl/kx5Q5i. K. Macek, “New report: Reversal in three-year uptick in pedestrian fatalities,” 2014, available online at http://goo.gl/D6UR6p. P. D. Loeb and W. A. Clarke, “The cell phone effect on pedestrian fatalities,” Transportation Research Part E: Logistics and Transportation Review, vol. 45, no. 1, pp. 284–290, 2009. A. H. Eichelberger and A. T. McCartt, “Toyota drivers’ experiences with dynamic radar cruise control, pre-collision system, and lanekeeping assist,” Journal of Safety Research, vol. 56, pp. 67–73, Feb. 2016. [Online]. Available: http://goo.gl/qvLMYF F. Seeliger, G. Weidl, D. Petrich, F. Naujoks, G. Breuel, A. Neukum, and K. Dietmayer, “Advisory warnings based on cooperative perception,” in Proceedings of IEEE Intelligent Vehicles Symposium Proceedings (IV), 2014, pp. 246–252. U.S. Department of Transportation, “Dedicated short range communication,” available online at http://goo.gl/MK1iDP. “World-first technology connects cycle helmets with cars,” 2014, available online at https://goo.gl/DgrblC. W. Hedgecock, M. Maroti, A. Ledeczi, P. Volgyesi, and R. Banalagay, “Accurate real-time relative localization using single-frequency GPS,” in Proceedings of ACM SenSys, 2014, pp. 206–220. T. Unger and V. Sandner, “The adac advanced emergency brake system test-a real life based approach for a better primary safety,” Berichte der Bundesanstalt fuer Strassenwesen. Unterreihe Fahrzeugtechnik, no. 87, 2013. T. Wang, G. Cardone, A. Corradi, L. Torresani, and A. T. Campbell, “WalkSafe: A pedestrian safety app for mobile phone users who walk and talk while crossing roads,” in Proceedings of ACM HotMobile, 2012, pp. 5:1–5:6. K. Dhondge, S. Song, B.-Y. Choi, and H. Park, “WiFi-Honk: Smartphone-based beacon stuffed wifi car2x-communication system for vulnerable road user safety,” in Proceedings of IEEE VTC2014-Spring, May 2014, pp. 1–5. S. Jain, C. Borgiattino, Y. Ren, M. Gruteser, Y. Chen, and C. F. Chiasserini, “Lookup: Enabling pedestrian safety services via shoe sensing,” in Proceedings of ACM MobiSys, 2015, pp. 257–271. C. Ericson, “Chapter 7 - spatial partitioning,” in Real-Time Collision Detection, ser. The Morgan Kaufmann Series in Interactive 3D Technology, C. Ericson, Ed. San Francisco: Morgan Kaufmann, 2005, pp. 285 – 347. [Online]. Available: http://goo.gl/bW9dDF Transport Accident Commission, “Pedestrian statistics,” 2014, available online at http://goo.gl/7Fij3e.