Professional Documents
Culture Documents
Abstract—Vehicle manoeuvre analysis is vital for road safety network (CNN) is a popular DL-based algorithm used for
as it helps understand driver behaviour, traffic flow, and road tasks such as image, text, and data classification and
conditions. However, classifying data from in-vehicle recognition [8].
acquisition systems or simulators for manoeuvre recognition is
complex, requiring data fusion and machine learning (ML) This research aims to classify vehicle manoeuvres data by
algorithms. This paper proposes a hybrid approach that extracting and fusing information from telematics. However,
combines multivariate multiscale entropy (MMSE) and one- the usual approaches for handling vehicle telematics data do
dimensional convolutional neural networks (1D-CNNs). MMSE not produce optimal outcomes. As a solution, advanced
is utilised for early feature extraction and data fusion, and the techniques are employed here to analyse manoeuvres by
extracted features are classified using 1D-CNNs, achieving an combining and extracting key features from telematics data
impressive 87% test accuracy in multiclass classification. This through statistical and machine learning methods. The MMSE
paper provides insights into improving vehicle manoeuvre technique extracts and fuses features, while the 1D-CNN
classification using advanced ML techniques and data fusion algorithm predicts outcomes. The remainder of this paper is
methods to handle complex data sets effectively. Ultimately, this organised as follows. Section 2 summarises works on
approach can enhance the understanding of driver behaviour, vehicular data, feature extraction, and fusion. Details of the
inform policy decisions, and develop more effective strategies to applied methodology are described in section 3. Section 4
enhance road safety.
presents the results with figures and discusses the performed
Keywords—Data Fusion, Data Extraction, Multivariate
analysis. Finally, section 5 concludes this paper with a
Multiscale Entropy (MMSE), Vehicle Manoeuvre. summary.
297
Authorized licensed use limited to: Malardalen University. Downloaded on March 28,2024 at 00:01:22 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE 13th International Conference on System Engineering and Technology (ICSET), 2 October 2023, Shah Alam, Malaysia
developed to identify and split the raw data into these classes. group is presented in Table 3. The same grouping process was
In the beginning, GPS data is manually observed to determine applied for each class.
the centre point of each turn, which is then inserted into a
program to identify all centre points in the raw data. The TABLE III. DIFFERENCE GROUPS BASED ON PARAMETERS.
distance between each centre point is then calculated to Group 1 Group 2 Group 3
determine simple and complex turns. A simple turn is 1. Yaw Rate Ext Sns 1. Vehicle Velocity 1. Pitch Ext Sns
recognised if the distance between the centre point to the 2. Speed Lateral 2. Speed Forward 2. Vert Accel Ext
3. Acceleration Lateral Sns
previous and next centre points is above a certain threshold. 4. Steering Wheel
Five values are then selected from that centre point's Angle
immediate left and right to construct the turn. More than two Group 4 Group 5
centre points can be involved in a complex turn. For example, 1. Pitch Rate Ext Sns 1. Acceleration
if the distance between centres A and B is below the 2. Vert Vel Ext Sns Forward
2. Acceleration Pedal
threshold, they will construct a complex turn together. The Position
turn is built using the five values from the left side of centre
A, five values from the right side of centre B, and all values 4) Phase 4: During this phase, the MMSE is used to
in the middle of both centre points. This research considered calculate entropy values for each group separately. The
three centre points to build a complex turn based on distance. MMSE includes the development and evaluation of
After identifying simple and complex turns, the straight data multivariate sample entropy (MSampEn) over multiple time
set was created using the remaining data. scales to perform analysis [5]. The MSE, which is a univariate
In the 45-minute driving time, straight, simple, and method proposed by Costa et al. [26], is used to measure the
complex turns are found in various positions along the complexity of a single channel. In contrast, the MMSE
trajectory. After identifying them, their frequency throughout
assesses the complexity of multichannel observations.
the 45 minutes is calculated and separated. Furthermore, the
sub-parts of straight and simple turns are marked serially for MMSE is performed in two steps where in the first step,
better clarity in Figure 1. Based on Figure 2, during phase 5, temporal scales are defined by averaging a -variate time
all samples are separated into bunches and undergo MMSE series , , = 1, 2, . . , over the non-coinciding time
to produce entropy values. Each bunch must contain at least segments of coarse granting, where is the number of
300 samples [5]. However, as the number of samples is low, samples in each channel. In this way for scale , coarse
the up-sampling method is employed to increase the sample gained multivariate time series is achieved as , =
size. During up-sampling, a certain number of random values ∑ , where 1 ≤ ≤ and the channel index is
are generated between two samples. For example, if two = 1, … , . In the second step, each coarse-grained
values are 2.5 and 3.4, a specific number of random values multivariate is evaluated by MSampEn and then plot
are generated within the range of 2.5 to 3.4. Samples for each
MSampEn as a function of the scale factor . As per the
instance of every activity were increased to 300 using this
approach. Table 2 displays the total number of samples reference cited in the article [5], it has been established that a
following the up-sampling process. minimum of 300 samples are required to estimate MSampEn
and a scale factor of eight is also taken into consideration in
TABLE II. LIST OF CLASSES AND NUMBER OF SAMPLES AFTER UP- this study. For further details regarding the computation and
SAMPLING.
algorithm of MMSE, refer to [5], [26]-[28].
Classes Sub-part Times of appearance Up-sampling
First 224 67200 5) Phase 5 and 6: During phase 5, the entropy values are
Second 223 66900 obtained by applying MMSE on the upsampled data of each
Simple Turn Third 223 66900
Fourth 222 66600 group. The upsampled data is split into 300-sample bunches
Fifth 214 64200 according to class wise, which are then fed to MMSE. As
Complex Turn - 216 64800 there are eight scale factors, MMSE provides eight entropy
First 216 64800 values for each bunch. For instance, group one has all its
Second 216 64800 classes' upsampled data, as shown in Table 2, with a total of
Third 216 64800 905400 samples. Since group one has four parameters, the
Fourth 216 64800
Straight
Fifth 225 67500 dimension of the dataset is 905400×4. After the data is fed
Sixth 204 61200 to MMSE, the final output is 3018×8, and this process is
Seventh 203 60900 repeated for the remaining four groups. In phase 6, the values
Eight 200 60000
extracted from each group are combined to form a final
dataset. The dimensions of the aggregated dataset are
3) Phase 3: Grouopwise data preprocessing, that is, 12
3018×40. A total of six datasets were prepared for
parameters mentioned in Phase 2 were used to create five
classification - five individual group-wise datasets and one
distinct groups. The first step was to analyse the correlation
aggregated dataset of five groups. The averaged MMSE
of each parameter. Based on their correlation score, the 12
features for each group in class-wise are displayed in Figure
parameters were divided into separate groups. For example,
3.
parameters such as "vehicle velocity" and "forward speed"
were grouped together due to their high correlation where the
correlation value is 0.97. The distribution of each parameter
298
Authorized licensed use limited to: Malardalen University. Downloaded on March 28,2024 at 00:01:22 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE 13th International Conference on System Engineering and Technology (ICSET), 2 October 2023, Shah Alam, Malaysia
dense_20 (None, 3) 51
Total param: 23,971
Trainable params: 23,971
(c) Group 3 (d) Group 4 Non-trainable params: 0
299
Authorized licensed use limited to: Malardalen University. Downloaded on March 28,2024 at 00:01:22 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE 13th International Conference on System Engineering and Technology (ICSET), 2 October 2023, Shah Alam, Malaysia
TABLE V. CLASSIFICATION ACCURACY RESULTS(%). another. The data was then subjected to binary classification
Group Data Split 1D-CNN ANN SVM KNN using 1D-CNN, ANN, SVM, and k-NN. The results revealed
Train 77.45 77.01 75.52 81.49 that 1D-CNN performed the best with an 83% test accuracy,
Group 1 Validation 78.14 77.15 73.67 74.36 surpassing the other models. Figure 6 shows the roc curve,
Test 74.56 75.66 75.16 73.00 which clearly indicates that 1D-CNN outperformed all the
Train 74.30 75.52 72.32 78.23 other models.
Group 2 Validation 71.02 70.86 70.69 65.23
Test 71.02 70.86 67.21 67.88 TABLE VII. RESULT FROM DECISION FUSION.
Train 65.41 64.53 63.42 71.60
Group 3 Validation 63.41 62.41 58.72 53.29 Prediction 1D-CNN ANN SVM k-NN
Test 62.08 59.60 60.63 56.78 326 291
Train 66.13 63.86 62.54 72.43 Correct 357 331
Group 4 Validation 61.75 61.58 62.41 60.43 278 313
Incorrect 247 270
Test 63.57 61.42 62.25 58.44
Train 65.85 65.30 63.25 71.10
Group 5 Validation 63.57 64.56 57.45 57.11
Test 63.41 62.08 64.40 56.29
Train 89.77 89.97 85.58 79.88
Aggerate
Validation 86.92 85.63 80.46 68.04
d
Test 87.74 83.88 82.94 70.86
B. Discussion
This research aims to fuse data using MMSE and then
extract features from MMSE used for classification to
Fig. 4: 1D-CNN’s training and validation loss. recognise vehicle manoeuvre. There are many ways to
perform fusion, like model agnostic and model-based. In the
model-based, researchers usually used kernel, graphical or
NN-based methods [6], [7]. This study utilised the MMSE, a
statistical method to merge multiple data sets and extract
noteworthy features. Extracted features get new
representation. Here, 12 parameters are grouped by
correlation, up-sampled, and then fed into MMSE for feature
extraction. Figure 3 presents the average of features of all
groups. From the figure, it is visible that the representation of
each group is different. Each group's classification result is
unsatisfactory, as presented in Table 5. The applied 1D-CNN
model achieved high accuracy for groups 2,3 and 4, but the
Fig. 5: Confusion matrix displaying the performance of 1D-CNN on the test test accuracy score was below 80%. The accuracy of other
data. models tested is also below 80%. Based on the group-wise
classification results, it can be concluded that individual
TABLE VI. F1 SCORES OF THE AGGREGATED GROUP ON THE TEST classification is not recommended because each group have
DATASET. less information which does not assist in increasing accuracy.
Group 1D-CNN ANN SVM k-NN
From Table 5, the classification result on the aggregated
Aggregated 86.49 79.65 77.83 64.12 dataset is good and except for k-NN, the test score of other
models stayed above 80%, whereas 1D-CNN has 87.74%
accuracy. This occurs because the aggregated dataset includes
This study utilises the decision fusion method to analyse data from all groups. After 82 epochs, the training of the 1D-
test results obtained from multiple ML models. The process CNN model for the aggregated dataset was stopped using the
involves combining the decisions made for each test sample early stopping technique to prevent overfitting. The results
in five groups from a model and calculating the average by show that the model achieved convergence early in the
dividing the sum by five. The average result of each sample is training process, and the early stopping mechanism prevented
then compared to the true label, and a vote is conducted. The overtraining by stopping the training process once the
number of correctly identified test samples is determined by validation performance plateaued. The confusion matrix of the
the results of the vote. In Table 7, the result of decision fusion test accuracy using 1D-CNN is depicted in Figure 5, and from
is presented where 1D-CNN has more correct values than the figure, simple and straight turns are mostly mismatched
others. To ensure the accuracy of the findings, the data set was with each other. Table 6 presents the F1 score of the
combined by merging samples of simple and complex turns aggregated dataset's test accuracy. A high score for the 1D-
into one category, while straight turns were classified as CNN model indicates that it can generalise well to completely
300
Authorized licensed use limited to: Malardalen University. Downloaded on March 28,2024 at 00:01:22 UTC from IEEE Xplore. Restrictions apply.
2023 IEEE 13th International Conference on System Engineering and Technology (ICSET), 2 October 2023, Shah Alam, Malaysia
new data. This suggests that the model can perform reliably in [8] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, "A survey of
real-world situations. convolutional neural networks: analysis, applications, and prospects,"
IEEE transactions on neural networks and learning systems, 2021.
To compare the results of multiclass classification, binary [9] Chahinez Ounoughi and Sadok Ben Yahia. Data fusion for its: A
classification is also conducted. The aggregated data is systematic literature review. Information Fusion, 2022.
transformed so that two classes become one and then [10] I. Cassias and A. L. Kun, "Vehicle telematics: a literature review,"
classified with another class using the same methods as Univ. New Hampshire, Durham, NH, USA, ECE. P, vol. 54, 2007.
before. The 1D-CNN method yielded the best results, [11] S. Kirushanth and B. Kabaso, "Telematics and road safety," in 2018
2nd International Conference on Telematics and Future Generation
achieving an accuracy of 83%. To further analyse the data, Networks (TAFGEN), 2018: IEEE, pp. 103-108.
decision fusion was performed based on the model's decisions
[12] D. A. Johnson and M. M. Trivedi, "Driving style recognition using a
of each group. Once again, 1D-CNN outperformed the other smartphone as a sensor platform," in 2011 14th International IEEE
methods, and the results can be found in Table 7. In addition, Conference on Intelligent Transportation Systems (ITSC), 2011: Ieee,
end-to-end neural network (NN) approaches are utilised for pp. 1609-1615.
data fusion and extraction. Autoencoder and variation [13] J. Paefgen, F. Kehr, Y. Zhai, and F. Michahelles, "Driving behavior
autoencoder (VAE) models are used and then classified using analysis with smartphones: insights from a controlled field study," in
CNN. However, the extracted feature is often inadequate and Proceedings of the 11th International Conference on mobile and
ubiquitous multimedia, 2012, pp. 1-8.
can result in overfitting during classification.
[14] K. Saleh, M. Hossny, and S. Nahavandi, "Driving behavior
classification based on sensor data fusion using LSTM recurrent neural
V. CONCLUSION networks," in 2017 IEEE 20th International Conference on Intelligent
This article presents a new method for identifying three Transportation Systems (ITSC), 2017: IEEE, pp. 1-6.
types of vehicle manoeuvres using various telematics data. [15] W. Weidner, F. W. Transchel, and R. Weidner, "Classification of scale-
The study analysed 12 telematics parameters and sensitive telematic observables for riskindividual pricing," European
Actuarial Journal, vol. 6, pp. 3-24, 2016.
preprocessed the data to distinguish between simple and
[16] Y. L. Murphey, R. Milton, and L. Kiliaris, "Driver's style classification
complex turns, as well as straight segments. MMSE was used using jerk analysis," in 2009 IEEE workshop on computational
to extract multivariate features, which were then classified intelligence in vehicles and vehicular systems, 2009: IEEE, pp. 23-28.
individually and combined into a whole group. Various ML [17] J. Xie, A. R. Hilal, and D. Kulić, "Driving maneuver classification: A
algorithms were employed for the classification task, comparison of feature extraction methods," IEEE Sensors Journal, vol.
including traditional methods like ANN, k-NN, SVM, and 18, no. 12, pp. 4777-4784, 2017.
1D-CNN. Experimental results showed that 1D-CNN was the [18] G. Singh, D. Bansal, and S. Sofat, "A smartphone based technique to
best method for binary and multi-class classification. monitor driving behavior using DTW and crowdsensing," Pervasive
and Mobile Computing, vol. 40, pp. 56-70, 2017.
However, there is still room for improvement, such as
incorporating data from other domains and optimising [19] C. Woo and D. Kuli´c, "Manoeuvre segmentation using smartphone
sensors, " In 2016 IEEE Intelligent Vehicles Symposium (IV), pp. 572–
parameters. In conclusion, this study demonstrates the 577. IEEE, 2016.
potential of telematics data for manoeuvre recognition and [20] E. Keogh and J. Lin, "Clustering of time-series subsequences is
lays the foundation for future research in this area. meaningless: implications for previous and future research,"
Knowledge and information systems, vol. 8, pp. 154-177, 2005.
ACKNOWLEDGMENT [21] Z. Li, L. Chen, J. Peng, and Y. Wu, "Automatic detection of driver
This work was supported in part by the project FitDrive fatigue using driving operation information for transportation safety,"
Sensors, vol. 17, no. 6, p. 1212, 2017.
(This project has received funding from the European Union’s
Horizon 2020 research and innovation programme under grant [22] C. Papadelis et al., "Monitoring sleepiness with on-board
electrophysiological recordings for preventing sleep-deprived traffic
agreement No 953432). accidents," Clinical Neurophysiology, vol. 118, no. 9, pp. 1906-1922,
2007.
REFERENCES [23] M. Siami, M. Naderpour, and J. Lu, "A mobile telematics pattern
[1] K. Wang, J. Yang, Z. Li, Y. Liu, J. Xue, and H. Liu, "Naturalistic recognition framework for driving behavior extraction," IEEE
Driving Scenario Recognition with Multimodal Data," in 2022 23rd Transactions on Intelligent Transportation Systems, vol. 22, no. 3, pp.
IEEE International Conference on Mobile Data Management (MDM), 1459-1472, 2020.
2022: IEEE, pp. 476-481. [24] [1] A. Cura, H. Küçük, E. Ergen, and İ. B. Öksüzoğlu, "Driver
[2] M. D. Pesé, A. Ganesan, and K. G. Shin, "Carlab: Framework for profiling using long short term memory (LSTM) and convolutional
vehicular data collection and processing," in Proceedings of the 2nd neural network (CNN) methods," IEEE Transactions on Intelligent
ACM International Workshop on Smart, Autonomous, and Connected Transportation Systems, vol. 22, no. 10, pp. 6572-6582, 2020.
Vehicular Systems and Services, 2017, pp. 43-48. [25] [1] S. Mohine, B. S. Bansod, R. Bhalla, and A. Basra, "Acoustic
[3] M. I. Silva and R. Henriques, "Finding manoeuvre motifs in vehicle modality based hybrid deep 1D CNN-BiLSTM algorithm for moving
telematics," Accident Analysis & Prevention, vol. 138, p. 105467, vehicle classification," IEEE Transactions on Intelligent
2020. Transportation Systems, vol. 23, no. 9, pp. 16206-16216, 2022.
[4] H. Suda, M. Natsui, and T. Hanyu, "Systematic intrusion detection [26] M. Costa, A. L. Goldberger, and C.-K. Peng, "Multiscale entropy
technique for an in-vehicle network based on time-series feature analysis of complex physiologic time series," Physical review letters,
extraction," in 2018 IEEE 48th International Symposium on Multiple- vol. 89, no. 6, p. 068102, 2002.
Valued Logic (ISMVL), 2018: IEEE, pp. 56-61. [27] M. U. Ahmed and D. P. Mandic, "Multivariate multiscale entropy
[5] M. U. Ahmed and D. P. Mandic, "Multivariate multiscale entropy: A analysis," IEEE signal processing letters, vol. 19, no. 2, pp. 91-94,
tool for complexity analysis of multichannel data," Physical Review E, 2011.
vol. 84, no. 6, p. 061918, 2011. [28] W. Li, X. Shen, Y. Li, and Z. Chen, "Improved multivariate multiscale
[6] T. Baltrušaitis, C. Ahuja, and L.-P. Morency, "Multimodal machine sample entropy and its application in multi-channel data," Chaos: An
learning: A survey and taxonomy," IEEE transactions on pattern Interdisciplinary Journal of Nonlinear Science, vol. 33, no. 6, 2023.
analysis and machine intelligence, vol. 41, no. 2, pp. 423-443, 2018. [29] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, "A survey of
[7] [A. Barua, M. U. Ahmed, and S. Begum, "A Systematic Literature convolutional neural networks: analysis, applications, and prospects,"
Review on Multimodal Machine Learning: Applications, Challenges, IEEE transactions on neural networks and learning systems, 2021
Gaps and Future Directions," IEEE Access, 2023.
301
Authorized licensed use limited to: Malardalen University. Downloaded on March 28,2024 at 00:01:22 UTC from IEEE Xplore. Restrictions apply.