Visualization Of Driving Behavior Using Deep Sparse Autoencoder HaiLong Liu1, Tadahiro Taniguchi2 , Tosiaki Takano2, Yusuke Tanaka3, Kazuhito Takenaka4 and Takashi Bando4 1. The Graduate School of Information Science and Engineering, Ritsumeikan University 3. Technical Research Division, Toyota InfoTechnology Center Co.,Ltd
2. The College of Information Science and Engineering, Ritsumeikan University 4. Corporate R&D Div.3, DENSO CORPORATION
INTRODUCTION
EXPERIMENT RESULTS VISUALIZATION USING DRIVING CUBE
Driving behavioral data Problems
DSAE
Driving behavioral data is high-dimensional time-series data The units and distributions in time space of each dimension are different It is not very intuitive for drivers to understand their driving behavior
Feature extraction
Visualization method
Deep sparse autoencoder
Driving cube Driving color map
PROPOSED METHOD Raw data 10 D
Normalization 10 D
SAE
Deep sparse autoencoder
Windowing
100 D
100 D
Driving cube
Binary linear SVM
RGB color space
Driving color map
80 D 50 D
30 D
F-measure of PCA, KPCA, SAE and DSAE for training set
0.8
F-measure
Raw data
F-measure
Safe
Driving behavioral data
Preprocessing
KPCA
Proposed method
Visualization Dangerous
PCA
0.6 0.4
10 D
0.6 0.4
0.2
0.2
0
0 Forward velocity PCA
3D
Turn right KPCA
Turn left SAE
F-measure of PCA, KPCA, SAE and DSAE for test set
0.8
Stop for obstacle
Forward velocity
detection
DSAE
PCA
Turn right
Turn left
KPCA
SAE
DSAE
Stop for obstacle detection
VISUALIZATION USING DRIVING COLOR MAP
Sparse autoencoder Error function
Windowing process
SAE
Circuit 1 (PCA)
Circuit 1 (KPCA)
Circuit 1 (SAE)
Circuit 6 (PCA)
Circuit 6 (KPCA)
Circuit 6 (SAE)
Questionnaire for 8 subjects
Course 2
BP* method
Circuit 6 (DSAE)
KPCA
Reconstruction layer
Hidden layer (features)
Please rank results of driving color maps using PCA, KPCA, SAE and DSAE from the viewpoint of distinguish ability of driving behavior
Ranking order
Course 1
Application
Circuit 1 (DSAE)
PCA
Subject experiment
DSAE
Input
1 2
*** *** ** *
Holm method ***:p<0.05 ***:p<0.01 ***:p<0.001
3 4
ANOVA: p= 1.72 × 10−5
Next SAE
PCA KPCA SAE DSAE
Differences in features using DSAE for different circuits
Visible layer * To make the SAE converge faster, after calculating the differential, we use one-dimension search to find the optimal value of the learning rate
Deep car watcher
EXPERIMENT EXPERIMENT PURPOSES
EXPERIMENT CONDITIONS
Methods for comparison Principal components analysis (PCA) Kernel principal components analysis (KPCA) Sparse autoencoder (SAE) Deep sparse autoencoder (DSAE)
Course 1
Course 2
Circuits 1~5
Circuits 6~10
Visualization using driving cube Using driving cube to compare the distribution of features Using binary support vector machines (SVM) with linear kernel for evaluation of methods. Because human can easily recognize patterns which can be linearly separable. Verifying generalization performances of methods Visualization using driving color map Using driving color map to compare the change of colors on map Determining the differences in features for different circuits Using questionnaire for evaluation of separability for driving color map
Training set
Test set
Circuits 1~4 & 6~9
Circuits 5 & 10
Frame rate
10 [fps]
Dimension
10 D
Accelerator opening rate [%] Meter of speed [km/h] Brake Master-Cylinder pressure [MPa] Steering angle [deg] Information Speed of wheels [km/h] of Engine speed [rpm], dimensions Longitudinal acceleration [m/s2] Transverse acceleration [m/s2] Yaw rate [rad/s] Turn signal [Right:-1,Left:1,OFF:0]
PCA
KPCA
SAE
DSAE
Window size
10
10
1
10
Dimensions variation
100→3
100→3
10→3
100→80→50 →30→10→3
Parameters
—
Gaussian kernel μ=0 σ2=1
α=0.15 β=0.7 θ=0.5
α=0.15 β=0.7 θ=0.5
People passing through pedestrian crossing
Braking due to obstacles after turning right
Without using turn signal
No one passing through pedestrian crossing
Not braking due to obstacles after turning right
Using turn signal
Stopping for people were passing through pedestrian crossing in front of car
Braking due to obstacles after turning right
Turn right on the corner of a right-hand turning lane without an intersection
CONCLUSION 1. DSAE was employed to extract the high-level low-dimensional features of driving behavior 2. Two visualization methods for high-level features were proposed, which were driving cube and driving color map. 3. Feature extraction using the DSAE is good for visualization from the viewpoint of separability of driving behaviors and has good generalization performance. 4. Driving cube and driving color map using the DSAE results in good visualization for time series data of driving behavior to determine the differences in such behavior. For 2014 IEEE Intelligent Vehicles Symposium June 8 - 11, 2014, Dearborn, Michigan, USA