Classification of Human's Driving Behavior using Support Vector Machine
1. Classification of Human’s Driving Behavior
Using Support Vector Machine
Graduate School of Information Science
Edahiro & Kato Laboratory
Yuki Kitsukawa
yuki@ertl.jp
1
RWDA 2015: Project Work
6. Objective
Hypothesis
It is possible to judge how to control vehicles based on learning models.
Verification Method – Support Vector Machine
1. Create learning model of surrounding environment and driving
behavior
2. Classify …
• Whether or not the driver steps the brake based on surrounding
environment
• If there are pedestrian around the vehicle based on driving
behavior
20. Conclusion
I researched the relationship between surrounding
environment and driving behavior through classification using
Support Vector Machine
Surrounding Environment → Driving Behavior
Whether to step break pedal: 88.3%
Driving Behavior → Surrounding Environment
Whether there is a pedestrian: 88.3%
21. Future Work
• Feature value
– Relative Position of pedestrian, vehicle
– Driving area (traffic environment, city, rural area…)
– Pedestrian’s direction
– Traffic Light
– Vehicle’s destination
– …
• Collect more dataset
• Parameter Tuning
Notes de l'éditeur
Today, I want to talk about how the surrounding environment around the vehicle affects the driver’s behavior.
The objective of this project is to invest whether there is a relation between surrounding environment and driver’s behavior especially I focused on the driver steps the brake pedal.
The objective of this project is to invest whether there is a relation between surrounding environment and driver’s behavior especially I focused on the driver steps the brake pedal.
Here, I will explain about the dataset. How I acquire the data. These are the sensors I used in this experiment. Grasshopper3 is the camera installed to capture the image of front of vehicle. This time the camera is used to detect pedestrians in front of the vehicle. Velodyne HDL-64E is the laser scanner installed on top of the vehicle to recognize the objects around the vehicle. This time velodyne is used to measure the distance from the vehicle to the pedestrian. CAN(Controller Area Network) signal is acquired through the CardBUS connected to the vehicle. From CAN signal, we can find the driver’s behavior. For example, the velocity of the car, how the driver step the accel, brake, how degree the driver turn the steering and so on. In this project, I combined the data acquired through these sensors.
To acquire the data, I conducted field operation experiment in imitation city in Toyota. Combining the image and velodyne data, we can estimate the distance to the pedestrian.
This is the dataset I acquired through experiment. Ispedestrian is the flag. If there are pedestrian captured by camera, it will be 1. Brakepress is the flag, if the driver steps the pedal, it will be 1.
This is the dataset I acquired through experiment. Ispedestrian is the flag. If there are pedestrian captured by camera, it will be 1. Brakepress is the flag, if the driver steps the pedal, it will be 1.
This is the dataset I acquired through experiment. Ispedestrian is the flag. If there are pedestrian captured by camera, it will be 1. Brakepress is the flag, if the driver steps the pedal, it will be 1.
Here, I want to conclude my project. I built the SVM classifier. The accuracy rate of pattern 1 is 77% and the pattern 2 is 56.7%.
I can say that accuracy rate of pattern 1 is relatively high. In other words, It can be predictable whether the driver step the brake pedal according to the surrounding environment.
However, it is difficult to estimate whether there is pedestrian from the CAN signal.
To improve the accuracy rate, it is necessary to improve the detection of pedestrian. The pedestrian detection program used in this experiment often makes miss-detection. There is a room to improve the detection. The second is, here I took into consideration whether there is a pedestrian or not, so it is good way to think pedestruan’s direction, for example, the pedestrian is walk along the road or the pedestrians is about to cross the road. Third is capturing other vehicles round itself. The other is adding sensors, backward camera or laser scanner and so on. And considering other algorithm for machine learning.