This slide was presented at the 2018 NOMS conference.

1. An IMU-based Turn Prediction System
Yu-Chang Ho1, Pei-Chen Lee2, Hsin-Yu Yeh2,
Min-Te Sun3, An-Kai Jeng4
1University of California, Davis
2Taiwan Semiconductor Manufacturing Company
3National Central University
4Industrial Technology Research Institute
Presenter: Yu-Chang Ho
2018.04.27 at NOMS 2018 Mini-conference Section
NOMS 2018

2. Outline
• Introduction
• Related Work
• System Design
• Field Experiment
• Conclusion
NOMS 2018 Outline 01 / 24

3. Introduction
• Many fatal car accidents are rear-ended collision.
• Major reason is human negligence, for example, forget to
use the turn signal.
• Many countries like the US and Taiwan, adopted new
technologies to address this issue.
• Many solutions required additional installation of sensors
or camera and might be affected by environmental
factors.
NOMS 2018 Introduction 02 / 24

4. Introduction(cont.)
• Nowadays, smartphones are widely available.
• We proposed a turn prediction system using the data by
the Inertial Measurement Units of driver's smartphone.
• We adopted digital maps to perform the prediction.
• We adopted the particle filter technique to predict the
vehicle's future location.
• Our system does not required additional hardware or
infrastructure installation.
NOMS 2018 Introduction 03 / 24

5. Related Work
• The related work in this context has two categories.
• Image-based Approach:
- Using digital camera
- May be affected by environmental factors
• Sensor-based Approach:
- Beyond IMU sensors
- IMU sensors
NOMS 2018 Related Work 04 / 24

6. Related Work(cont.)
• Beyond IMU Sensors:
- Requires additional sensors, like laser ranging sensors or
ultrasonic sensors
- Specially designed sensors to detect the turning degree
• IMU Sensors:
- Analyze the state of acceleration/deceleration
- Predict driver's behaviors
NOMS 2018 Related Work 05 / 24

7. System Design
• The system contains three modules:
Data Collection:
- Collect data from smartphone's IMU sensors.
- Adopt moving average filters and complementary filters for noise
removal.
NOMS 2018 System Design 06 / 24

8. System Design(cont.)
Internal Calculation:
- In this module, the particle filter is adopted.
- Consists of five stages:
NOMS 2018 System Design 07 / 24
Initialization
System
State Prediction
Weights Update Resampling Expectation

9. System Design(cont.)
Internal Calculation - Initialization
- The particles are normally distributed in the IMU
sensor errors range
NOMS 2018 System Design 08 / 24

10. System Design(cont.)
Internal Calculation - System State Prediction
NOMS 2018 System Design 09 / 24
- Each particle will obtain its
speed and direction info. from
the IMU sensors.
- Compute the distribution of the
particles for the next time slot.
- The average position will be the
predicted location of the
vehicle.

11. System Design(cont.)
Internal Calculation - Weights Update
NOMS 2018 System Design 10 / 24
- Particles that are outside the
GPS error range are removed.
- The weights of the remaining
particles are updated.

12. System Design(cont.)
Internal Calculation - Resampling
NOMS 2018 System Design 11 / 24
- Particles with heavier weights
are duplicated to refill.
- Initialize new particles speed
and direction info. based on
IMU data.

13. System Design(cont.)
Internal Calculation - Expectation
NOMS 2018 System Design 12 / 24
- Previous four stages will repeat
for several times.
- Now the system got the
historical, current, and
predicted position.
- Compute the curvature.

14. System Design(cont.)
Turn Prediction:
NOMS 2018 System Design 13 / 24
- Calculate the predicted route
curvature.
- Compare the predicted route
curvature with the curvature
of the road.
- The difference of curvatures
is denoted as Diffc.
- Tc: The threshold of Diffc
- Ts: The threshold of vehicle
speed

15. System Design(cont.)
• The flow chart of the turn prediction algorithm:
NOMS 2018 System Design 14 / 24

16. Field Experiment
• Routes: Hsinchu Government (HCHG), National Chiao
Tung University (NCTU), National Tsing Hua University
(NTHU)
NOMS 2018 Field Experiment 15 / 24

17. Field Experiment(cont.)
• To make sure the GPS is accurate enough, we applied the
High-precision GPS device made by ITRI.
• Compare the curvature obtained by Particle Filter,
smartphone GPS, and ITRI High-precision GPS.
NOMS 2018 Field Experiment 16 / 24

18. Field Experiment(cont.)
• Confusion Matrix:
• True Negative (TN): the system predicts not make a turn,
and the vehicle is not.
• TN is not considered by our system.
NOMS 2018 Field Experiment 17 / 24

19. Field Experiment(cont.)
• Accuracy:
• Error rate:
NOMS 2018 Field Experiment 18 / 24

20. Field Experiment(cont.)
• Determine Tc using several tests.
• Determine Ts:
NOMS 2018 Field Experiment 19 / 24

21. Field Experiment(cont.)
• The accuracy and error rate for routes in HCHG.
NOMS 2018 Field Experiment 20 / 24

22. Field Experiment(cont.)
• Prediction Time Difference:
- ∆T1: The time difference using smartphone GPS
- ∆T2: The time difference using high precision GPS
NOMS 2018 Field Experiment 21 / 24
detection time

23. Field Experiment(cont.)
NOMS 2018 Field Experiment 22 / 24
Left-turn Timing Dataset

24. Field Experiment(cont.)
• Average Prediction Time Difference:
• Our results is 0.197 seconds, assume 40 km/hr is the
average vehicle speed, the particle filter could detect a
turn event earlier by 2.19 meter.
NOMS 2018 Field Experiment 23 / 24

25. Conclusion
• We proposed a light-weight turn prediction system.
• The proposed system does not require additional
hardware or infrastructure.
• The conducted field experiments shown that our system
works in a real-world environment.
• The design of the application part could be the future
work of this research.
NOMS 2018 Conclusion 24 / 24

26. Thank you!
2018.04.27 NOMS 2018 Yu-Chang Ho