Method of Extracting Characteristics of Truck Road Usage from GPS Data

Size: px

Start display at page:

Download "Method of Extracting Characteristics of Truck Road Usage from GPS Data"

Valerie Sanders
5 years ago
Views:

1 Method of Extracting Characteristics of Truck Road Usage from GPS Data Takayoshi Yokota Advanced Transport Logistics (Hanshin Expressway) Department of Urban Management Graduate School of Engineering Kyoto University 1

2 Outline 1 Objectives 2 Overview of truck probe experiment 3 Concept of two layered road network for trucks 4 Model of the layered network 5 Characteristics of the layered network 6 Route identification from GPS probe data based on DP 7 Conclusions 2

3 1. Objectives Understand, model and evaluate the behavior of goods vehicles from a macroscopic viewpoint. As the macroscopic viewpoint spatial distribution of trucks is considered. To examine the goods vehicle behavior, we conducted probe car experiment in October,

4 2. Truck Probe Experiment Period : October 1 to 31, 2009 Participants: 300 trucks belonging to 21 different companies Area: Keihanshin Area (2 nd largest urban area in Japan, population 19 million, consisting of Kyoto, Osaka and Kobe) 4

5 Experiment Overview Carrier Base Data Acquisition Time Stamp Latitude, Longitude Data Acquisition Server GPS Data Analyzer GPS Cell phone Trucks 7

6 Acquired probe data and its density 8

7 Number of trucks Daily total mileage Acquired probe data Number of trucks Daily total mileage Date Date Weekdays: 270 trucks Weekdays: 40,000 km 9

8 2009/10/13 A 6 vehicles B 20 vehicles C 20 vehicles D 30vehicles E 15vehicles F 5 vehicles 12

9 S 18 vehicles T 6 vehicles U 5 vehicles 2009/10/13 V 2 vehicles If they are put together they seem to cover the Keihanshin Area 15

10 3. Concept of two layered road network for trucks High Frequency Trucks prefer major roads such as motorways to narrow roads, because they concern safety, reliability of arrival time. Op Dq 2 layered network model is proposed access egress op Low Frequency dq Simplify and improve the route choice model by reducing the size of choice set to be more realistic. Computation power consumption will decrease. 16

11 4.1 Model of the layered network Probe Density Map The number of probe data which fall into each cell is counted for one month. Major roads and minor roads are clearly separated. They look like human artery(or vein ) and capillary 20m 20m Cell The size of cell is determined taking GPS error into account. 17

12 >76 probes/cell day Cells which satisfy the threshold 18

13 >67 probes/cell day Cells which satisfy the threshold 19

14 >57 probes/cell day Cells which satisfy the threshold 20

15 >38 probes/cell day Cells which satisfy the threshold 21

16 >28 probes/cell day Cells which satisfy the threshold 22

17 >19 probes/cell day Cells which satisfy the threshold 23

18 >14 probes/cell day Cells which satisfy the threshold 24

19 >9.5 probes/cell day Cells which satisfy the threshold 25

20 >4.7 probes/cell day Cells which satisfy the threshold 26

21 >2.6 probes/cell day Cells which satisfy the threshold 27

22 >1.3 probes/cell day Cells which satisfy the threshold Sudden Increase! 28

23 >0.66 probes/cell day Cells which satisfy the threshold Sudden Increase! 29

24 すべて >0.19 probes/cell day Cells which satisfy the threshold Sudden Increase! 30

25 >0.15 probes/cell day Cells which satisfy the threshold Sudden Increase! 31

26 >0.03 probes/cell day Cells which satisfy the threshold Sudden Increase! 32

27 Number of cells Number of cells Number of cells 4.2 Model of the layered network The distribution of the number of cells with respect to the minimum probe density can be approximated as a sum of two different exponential distributions. Minimum probe density (counts/cell month) High Density(Frequency) Low Density(Frequency) + Minimum probe density (counts/cell month) Minimum probe density (counts/cell month) 33

28 4.3 Number of cells with respect to the minimum probe density R 2 = times faster Close Up 34

29 4.4 Number of cells with respect to the minimum probe density 35

30 4.5 Number of cells with respect to the probe density K(q):The number of cells with respect to the probe density is defined as: K is given by: 36

31 4.7 Number of cells with respect to the probe density (log scale) This model also fits well. 38

32 4.8 Extracted two layered network Extracted Network Digital Roadmap (width >5.5m ) You can neglect excessive roads. High Frequency Network Low Frequency Network 39

33 4.9 Extracted two layered network Two Layered Network High Frequency Network consists of at least Motorways and National roads. Digital road map High frequency network Low frequency network Motorway Motorway National Road 40

34 5.1 Characteristics of the network - Distribution of Probes Threshold is chosen to 30 counts/cell*month. 41

35 5.2 Characteristics of each network Road length and running time Number of Cells( Road Length) Low Frequency Network 13% Running Time High Frequency Network 26% Low Frequency Network 74% High Frequency Network 87% 42

36 5.3 Characteristics of the network: Comparison of instantaneous velocity 43

37 5.4 Characteristics of the network: Comparison of instantaneous velocity Mean:25.0km/h Mean:16.9km/h HF layer : Velocity shifted to higher speed 44

38 Summary : Road networks are separated into 2 layers Keihanshin Road Network Separated into 2 layers From the Frequency Viewpoint High Frequency Network 1)Average 25km/h 2)Used 6 Ttmes longer time 3) Size 1/4 Low Frequency Network 1)Average 17km/h 45

39 Road Usage with respect to class of road Trip length (km) Trip length (km) Frequency of using motorways(hanshin expressway, NEXCO highways) decreases as the length of trips decrease. 2015/6/3 46

40 6. Route identification from GPS probe data It is not easy to identify routes only from GPS data due to the following obstacles. 1)GPS data suffers roughly 30m error. 2)Map data suffers roughly 20m error. 3)Many elevated motorway with road running just beneath in urban area. e.g. Hanshin Expressway Route 5 and National route 43 Hanshin Expressway Route 13 and National route

41 6. Route identification from GPS probe data If we only know present and past we cannot guess the route properly. future present past Motorway Usual road If we can also know future, we can guess the answer quite easily. future present past Motorway Usual road 48

42 6. Route identification from GPS probe data speed link l j Using speed is difficult. This may need site specific DBs. node a node a Previous probe point P p (t i-1 ) Probe point P p (t i ) Vehicle heading Current probe point P p (t i ) Link heading Interpolation point node b Interpolation point node b Projection point P m (t i ) Projection point P m (t i ) Interpolation point Interpolation point Interpolation point Interpolation point distance heading 49

43 6. Route identification from GPS probe data distance heading Speed(controversial) Find a series of discrete decisions s(i),i=0,..,p-1 which maximizes Recursive formulation Dynamic Programming 50

44 6. Route identification of from probe data based on DP(Dynamic Programming) E(0,0) E(0,1) E(0,j ) E(0,N-1) Stage 0 + f(0,0) + f(0,1) + f(0,j ) + f(0,n-1) Connection matrix : a(1,j,j)=1, if reachable a(1,j,j)=0, otherwise E(1,0) max {f(0,j ) a(1,j,j)}} j E(1,1) max {f(1,j )} a(1,j,j)} j E(1,j) max {f(1,j ) a(1,j,j)} j E(1,N-1) max {f(1,n-1) a(1,j,n-1)} j Stage f(1,0) f(1,1) f(1,j) f(1,n-1) Connection matrix : To higher stages 51

45 6.Route identification probe data based on DP(Dynamic Programming) EXAMPLE Profile of DP score Result of route identification Profile of class of road e.g. vehicle no.2, Hanshin expressway These spikes can be eliminated. 52

46 6.Route identification probe data based on DP(Dynamic Programming) Result of route identification Profile of class of road 53

47 6.Route identification probe data based on DP(Dynamic Programming) 8:54-9:08 10:16-10:39 16:06-16:39 Result of route identification Profile of class of road 54

6.Route identification probe data based on DP(Dynamic Programming) Introducing speed sometimes brings side effects. Miss-match to parallel road due to the slow down on the expressway.

48 6.Route identification probe data based on DP(Dynamic Programming) Introducing speed sometimes brings side effects. Miss-match to parallel road due to the slow down on the expressway. Miss-match to parallel expressway due to the increased speed on arterial road. Introduced speed factor This doesn t help much Minimum duration of highway Slowdown on expressway Caused miss-match to a National road 5min 55

49 Detected On Ramp 56

50 57

51 Detected Off Ramp 58

52 59

53 Detected On Ramp 60

54 61

55 Detected Off Ramp 62

56 63

57 Detected On Ramp 64

58 65

59 Detected Off Ramp 66

60 67

61 7. Conclusions 1) The road network for trucks can be modeled as a two layered network, and both layers could be adaptively constructed according to the frequency of road use, which is calculated with probe data. 2) Trucks in the Keihanshin area spent about 6 times longer on the high frequency network even though it was 1/4 the size of the low frequency network. Thus, improving the reliability, efficiency, and safety of the high frequency network deserves a higher priority. 3) The instantaneous velocity histograms did reveal a significant difference between the two layers as one of their main characteristics, e.g., the mean value was km/h in the low frequency network and km/h in the high frequency network. 4) Evaluation of services, analysis of reliability and modeling of route selection should become easier to deal with and become simpler and more realistic if based on this two-layered network. (Currently under study.) 5) Dynamic Programming based algorithm for route identification is developed which can estimate accurate enough routes even from noisy GPS probe. 68

62 湾岸線, 環状線他利用 80 号車高速道路利用せず 83 号車高速道路利用せず 84 号車神戸線他利用 85 号車 ( 片道 )

63 阪神高速利用トラックの運転時間は最低でも 600 秒 (10 分 ) 以上は使っている

64 高速道路利用せず 66 号車高速道路利用せず 72 号車湾岸線利用 67 号車高速道路利用せず 74 号車高速道路利用せず 76 号車高速道路利用せず 78 号車湾岸線利用 79 号車

65 高速道路利用せず 88 号車高速道路利用せず 90 号車高速道路利用せず 91 号車

66 高速道路利用せず 88 号車高速道路利用せず 90 号車高速道路利用せず 91 号車

67 高速道路利用せず 95 号車高速道路利用せず 106 号車

68 ここは厳しい

69 101 号車 10/1 道頓堀付近 76

70 77

71 Actual number of trucks for each goods (vehicles/day) Actual number of trucks for each area (Vehicles/day) Chemical industrial goods 24 Inside of Osaka City 136 Machinery and metal industrial goods 30 North of Osaka Prefecture 25 Light industrial goods 196 East of Osaka Prefecture 35 Others 50 South of Osaka Prefecture 60 Outside of Osaka Prefecture 44 78

72 79

73 80

74 Heuristic Modification Post Processing Filter 1)Eliminate false data with get off and quick get on 2)Eliminate false data with get on and quick get off Class 0,3-9 Studied from 109 vehicles data Class 1,2 Parameters to accept class changes T ON_OFF T ON_OFF_ON T ON_OFF > 300 sec.. more than 99.5% tile Class 0,3-9 Class 1,2 T OFF_ON T ON_OFF_ON > 600 sec...more than 99.5% tile T * OFF_ON > 0 sec. There is nothing we can do about this. This case occurs when the trip starts initially from arterial roads (Class 0,3-9), which is almost always the case. 81

75 Heuristic Modification Post Processing Filter to increase accuracy of detecting on and offs of motorways Result of route identification Profile of class of road T ON_OFF > 300 sec % tile T ON_OFF_ON > 600 sec % tile 82

アルゴリズムの設計と解析. 教授 : 黄潤和 (W4022) SA: 広野史明 (A4/A8)

アルゴリズムの設計と解析. 教授 : 黄潤和 (W4022) SA: 広野史明 (A4/A8) アルゴリズムの設計と解析教授 : 黄潤和 (W4022) rhuang@hosei.ac.jp SA: 広野史明 (A4/A8) fumiaki.hirono.5k@stu.hosei.ac.jp Divide and Conquer Dynamic Programming L3. 動的計画法 Dynamic Programming What is dynamic programming? Dynamic