An Evaluation Study of Driver Profiling Fuzzy Algorithms using Smartphones German Castignani, Raphaël Frank, Thomas Engel Interdisciplinary Centre for Security Reliability and Trust (SnT) University of Luxembourg raphael.frank@uni.lu 1
Outline 1. Driver Risk Monitoring 2. Smartphone Sensors for Driver Monitoring 3. Driver Profiling Algorithms 4. System Architecture 5. Evaluation 6. Conclusion and Perspectives 2
Driver Risk Monitoring: Why? Fleet Management Pay-As-You-Drive Usage-Based Insurance Eco-Driving Driver Education 3
Measuring Driver s Risk (1/2) 80% of accidents are caused by inattention/distraction [1,2] Inattention-related crashes are four times higher young drivers (18-20) than for experienced drivers (>35) [1,2] Unsafe drivers perform hard acceleration, braking and steering maneuvers more frequently than safe and moderate drivers [1,3] Lateral Acceleration (G-Force) Longitudinal Acceleration (G-Force) [1] NHTSA, "The 100-Car Naturalistic Driving Study - Results of the 100-Car Field Experiment, DOT HS 810 593, 2006 [2] Ranney et Al., NHTSA Driver Distraction Research: Past, Present and Future, 2000 [3] NHTSA, An Exploration of Vehicle-Based Monitoring of Novice Teen Drivers: Final Report, DOT HS 811 333, 2010 4
Measuring Driver s Risk (2/2) Existing NHTSA research studies on the effect of monitoring young drivers [1,3] Hard braking and turning maneuvers also appear to be linked to increased crash/near-crash involvement. Therefore, if effective feedback can be provided to teen drivers that results in a decrease in the prevalence of these behaviors, their involvement in crashes and near-crashes may also be reduced. Teen driver monitoring systems should attempt to include some method of monitoring these behaviors and provide effective feedback regarding these behaviors to help reduce teen involvement in crashes and near-crashes. Monitoring can reduce accidents by 20% [4] Driving safer means driving smoother à Toyota Study estimates that the fuel consumption can be reduced by 10-30% [5] In the US, it is estimated that a mass adoption of insurance telematics could results up to $20 billion savings for insurance companies [6] [4] Wouters at al., Traffic Accident Reduction by monitoring driver behavior with in-car recorders, SWOV Institute for Road Safety Research, 2000 [5] Toyota, Numerous Benefits of Eco Driving, 2009 [5] Blase et al., Cars and Trucks are Talking: Why Insurers Should Listen, Diamond Management & Consultants, 2008 5
Existing solutions: Black Boxes Telematics boxes installed in individual cars Dedicated data communication interface Multiple embedded sensor (GPS, acceleration, CAN bus data ) Detect multiple driving maneuvers à Very high installation and maintenance cost à Limited feedback to drivers à Limited deployment 6
What about Smartphones for monitoring? Advantages: High penetration rate worldwide (USA 61%) [7] No additional cost for the monitoring entity Real tracking not required à Computation can be done locally Trust relationship: more trusted than a black-box Only limited amount of data to report Available SDKs and libraries à APP development They have all the required sensors! Forecast: 75% Sensor Penetration in Smartphones for 2017 [8] [7] Nielsen., MOBILE MAJORITY: U.S. SMARTPHONE OWNERSHIP TOPS 60%, 2013 [8] Yole Développement, The Growth of MEMS Market, September 2012 7
Smartphone integrated sensors Accel / Jerk Steering / Yaw-rate Speed Inattention Proximity to vehicle Environment Accelerometer " " Magnetometer " GPS " " " Camera(s) " " " Light Sensor " Barometer " Microphone " " Touchscreen " Wireless " " " " 8
Existing solutions using smartphones Insurance Companies: AVIVA UK: RateMyDrive StateFarm US: DriverFeedback Ingenie: Young Driver Car Insurance Startups: MOVELO à Spin-off negotiating with large Scandinavian insurance company SageQuest: Driver Style à Fleet Management Others: GreenRoad & Optimium à Eco Driving ionroad Personal Driving Assistant (ADAS like) MotorMate Driver Education (Governmental) 9
Driver Profiling Algorithms Eren et al. [9] / Johnson et al. [10] Driving event detector using Dynamic Time Warping (DTW) Paefgen et al. [11] Event detection using fixed thresholds Validation against CAN-bus data You et al. [12] Fuses GPS, sensors, front and rear cameras Drowsiness events detected with 85% accuracy Araujo et al. [13] Combines smartphone sensors and CAN-bus data Give advices to drivers to improve eco-efficiency [9] H. Eren et al., Estimating driving behavior by a smartphone, IEEE Intelligent Vehicles Symposium, 2012 [10] D. A. Johnson et al., Driving style recognition using a smartphone as a sensor platform, IEEE ITSC, 2011 [11] J. Paefgen et al., Driving behavior analysis with smartphones: Insights from a controlled field study, ACM Press, 2012 [12] C.-W.You et al., CarSafe: a driver safety app that detects dangerous driving behavior using dual-cameras on smartphones, ACM UbiComp, 2012 [13] R. Araujo et al., Driving coach: A smartphone application to evaluate driving efficient patterns, IEEE Intelligent Vehicles Symposium, 2012 10
Building Blocks 1. Gather data from sensors/gps and send it when the trip ends 2. Analyze data and compute input variables 3. Assign a score based on input data from different trips 11
Validation of Sensing data Test Conditions: Driver accelerates 150m and then steers (right and left) for 50m Two tests: Moderate and aggressive acceleration/steering Speed (m/s) 0 5 10 15 20 GPS Bearing Rate (degrees/s) 0 10 20 30 40 50 60 Total Sensor Acceleration (m/s2) 0 2 4 6 8 10 12 Raw data Kalman filter Low total accel 5 10 15 20 25 30 35 5 10 15 20 25 30 35 10 15 20 25 30 35 40 Time (s) Time (s) Time (s) (a) Experiment 1: Speed (b) Experiment 1: Bearing rate (c) Experiment 1: Sensor acceleration Speed (m/s) 0 5 10 15 20 Higher speed and GPS accel GPS Bearing Rate (degrees/s) 0 10 20 30 40 50 60 +60% bearing rate Total Sensor Acceleration (m/s2) 0 2 4 6 8 10 12 Raw data Kalman filter Aggressive accel and steering 5 10 15 20 25 30 35 5 10 15 20 25 30 35 10 15 20 25 30 35 Time (s) Time (s) Time (s) (d) Experiment 2: Speed (e) Experiment 2: Bearing rate (f) Experiment 2: Sensor acceleration Fig. 2. Sensing data from a single trip 12
Input Variables Sensor Acceleration Total acceleration as the Euclidean sum of the filtered acceleration vector. Count moderate (SA M, >1.5m/s 2 ) and aggressive (SA A, > 3m/s 2 ) acceleration events per kilometer [1]. GPS acceleration Speed variation (in m/s 2 ) Count moderate acceleration events (GA M, >1m/s 2 ) and aggressive events (GA A, 2.5m/s 2 ) Compute maximum acceleration (GA P ) and deceleration (GA N ) GPS steering/bearing Azimuth (direction to the north) variation (in /s) Count moderate steering events (BR P, >10 /s) and aggressive events (BR A, 40 /s) per kilometer. Over-speed Compute over-speed relative time (OS R ), maximum overs-peed and average overspeed (OS A ) 13
Fuzzy System 1 Fuzzy System: Allows merging heterogeneous/fuzzy data to come out with a single score Convert input data in linguistic variables LOW (L), MEDIUM (M), HIGH (H) Apply IF THEN rules: (example) IF GA P =H & GA N =H à AGGRESSIVE IF OS T =L & OS A =L & OS P =L à NORMAL 0.75 0.5 0.25 0 0 12.5 25 37.5 50 62.5 75 100 CALM AVERAGE MODERATE AGGRESSIVE 14
N centroid defuzzification method [16] to obtain the score (between 0 and 100). The centroid approach computes the center of gravity of the obtained output fuzzy curve. This score is then used to rank the different drivers, being 0 the best and 100 the worst possible score. In our problem, for each driver, we set up a Fuzzy Inference problem for each particular driving environment, i.e., urban (S u ), suburban (S s ) and extra-urban (S e ). Then, the final score is calculated using the weighting function in Equation 1. In particular we assign a greater weight to the urban score (i.e., w u =0.5) since we consider that aggressive behaviors are more risky in urban environments than in suburban (w s =0.3) and extraurban (w e =0.2) environments [17], due for example to the presence of pedestrians in urban environments. Final Score: combination of Urban, Suburban and Extra-urban Defuzzification and Scoring Rules trigger multiple categories with different membership levels S = w u S u + w s S s + w e S e (1) A. Testbed V. EVALUATION 0.50 0.30 0.20 In order to evaluate our profiling mechanism, we have dis- 15
Evaluation: Testbed Goal: Calculate a driving score for participants Test Conditions: Luxembourg area Corporate VW Caddy / Renault Kangoo Fleet Remote data collection from smartphones (3G/ WiFi) Total participants: 20 Aggregated path length: 8.600 km Total number of trips: 566 16
Evaluation: Results 17
Evaluation: Obtained Score Driver 4 18 2 9 7 3 10 12 19 8 17 14 6 5 13 15 1 11 16 20 45.15 46.01 48.36 48.49 48.49 50.48 52.47 52.97 55.26 58.96 65.82 66.87 68.82 69.31 70.21 71.76 71.76 74.07 75.78 77.65 0 20 40 60 80 100 Score 18
Conclusions and Perspectives We proposed a fuzzy system for driver profiling that takes as input variable sensing data from Android smartphones Limitation with sensors: Noisy accelerometer and magnetometer signals Difficulty to decouple longitudinal and lateral acceleration Consider Moderate and Aggressive acceleration/steering events after filtering A Gamification approach Scoring as a contest between anonymous users Sharing scoring data in a Social-Network approach 19
Ongoing work Event Detector Real-time event detection Fuzzy-Logic, SVM, ANN classification 20
Thank you for your attention! Questions 21