27th International Conference on Information Technology (IT)
Žabljak, 15 – 18 February, 2023
Vehicle Speed Estimation From Audio Signals
Using 1D Convolutional Neural Networks
Ivana Čavor and Slobodan Djukanović
Abstract— This paper presents an approach to acoustic
vehicle speed estimation using audio data obtained from single-
sensor measurements. One-dimensional convolutional neural
network (1D CNN) is used to estimate the vehicle’s speed
directly from raw audio signal. The proposed approach does not
require manual feature extraction and can be trained directly
on unprocessed time-domain signals. The VS13 dataset, which
contains 400 audio-video recordings of 13 different vehicles,
is used for training and testing of the proposed model. Two
training procedures have been evaluated and tested, one based
on determining optimal number of training epochs and the
other based on recording model state with minimal validation
loss. The experimental results show that the average estimation
error on VS13 is 9.50 km/h and 8.88 km/h, respectively.
I. I NTRODUCTION
Intelligent transport systems have become a crucial part
of the concept of smart cities, therefore the relevance of
automatic traffic monitoring (TM) systems has increased
considerably [1]. Different types of sensors enable the col-
lection of various data (text, sound, image, video) about
traffic conditions and their real-time processing. Information
regarding the number of vehicles on the roads, vehicle types,
estimated speed and acceleration of vehicles, are particularly
useful for the enforcement of legal regulations for speed
limits and traffic safety improvement. For example, in the
vicinity of speed cameras, the number of speeding vehicles
and collisions was reduced by 35% and 25% respectively
[2].
Modern TM systems are mostly based on computer vision
and enable the detection, tracking and classification of vehi-
cles based on digital images [3]–[6]. However, despite their
good performance, systems based on specialized cameras
are more expensive, complex to install and maintain, and
memory-intensive in terms of data processing and storage.
Acoustic sensors are becoming a viable alternative for visual
surveillance systems since they require significantly fewer
resources. In this paper, we estimate the vehicle’s speed
using the sound it makes as it passes by acoustic sensor
(microphone).
In recent years, various approaches for TM have been
introduced. The spectral and spatial characteristics of audio
signals are used in [7] for vehicle speed estimation and
Ivana Čavor is with the Faculty of Maritime Studies Kotor, University of
Montenegro, Kotor, Montenegro (e-mail: ivana.ca@ucg.ac.me)
Slobodan Djukanović with the Faculty of Electrical Engineering, Univer-
sity of Montenegro, Podgorica, Montenegro (e-mail: slobdj@ucg.ac.me)
979-8-3503-9751-2/23/$31.00 ©2023 IEEE
the authors claim that in single-microphone systems, signal
frequency information based on the Doppler effect is not
useful for estimating the vehicle’s speed. In [8], the speed
and position of vehicle are determined by analyzing the
characteristic changes in the spectrogram. The authors of
[9] consider several representations of audio signals and
use 85 discriminative features to identify speed variations
using different machine learning techniques. To estimate the
vehicle’s speed in [10], time-frequency characteristics of the
sound signal produced by the vehicle engine are used as input
to the neural network.
A new speed-dependent analytical function, referred to as
modified attenuation (MA) is proposed in [11]. Deep neural
network is used to estimate the MA feature based on the log-
mel spectrogram of the vehicle sound and support vector
regression is used to predict the vehicle’s speed. The MA
feature is used in [12] to correct the exact moment when the
vehicle passed by the sound sensor, resulting in a drop of the
mean error of the estimated vehicle speed from 7.39 km/h to
6.92 km/h. Mel-frequency representations (mel-spectrogram,
log-mel spectrogram, log-mel cepstral spectrogram) of the
sound signal produced by vehicle and a fully connected neu-
ral network are combined in [13]. It should be emphasized
that each mentioned approach employs some audio signal
processing technique before speed calculation.
In this paper, we propose vehicle speed estimation model
based on 1D CNNs, which can automatically extract impor-
tant features from raw audio signals. A brief introduction to
1D CNN is provided in Section II, whereas the proposed
model is described in Section III. Section III also briefly
describes the VS13 dataset of audio-video recordings used
for training and testing of the proposed model. Experimental
results are presented in Section IV. Finally, Section V gives
the conclusions and directions of the future research.
II. O NE - DIMENSIONAL C ONVOLUTIONAL N EURAL
N ETWORKS
A. 1D CNN overview
Deep learning models are developed based on the structure
and information processing system of the human brain.
While two-dimensional CNNs have found widespread use
in computer vision tasks, 1D CNNs are primarily intended
for processing one-dimensional sequences [14].
CNN is a specific type of deep neural networks that
use convolution layers, which are made up of a collection
of convolutional filters or kernels. Typically, convolutional
 TABLE I
T HE PROPOSED 1D CNN ARCHITECTURE

# Kernel Pool Output

Layer Stride
Filters Size Size Shape
Input – – – – 264600
Conv1D 32 32 – 16 32 × 16536
Conv1D 32 8 – 4 32 × 4133
AveragePool – – 2 2 32 × 2066
Conv1D 64 2 – 2 64 × 1033
Conv1D 64 2 – 2 64 × 516
AveragePool – – 2 2 64 × 258
Conv1D 96 2 – 2 96 × 129
Conv1D 96 2 – 2 96 × 64
AveragePool – – 2 2 96 × 32
Conv1D 128 2 – 2 128 × 16
Conv1D 128 2 – 2 128 × 8
Flatten – – – – 1024
Dense – – – – 100
Dense – – – – 10
Output – – – – 1

Fig. 1. Illustration of the typical structure of a 1D CNN model.
filters are applied over a number of layers, each of which ex-
tracting characteristics of the considered signal at a particular
abstraction level. CNNs enable reliable and effective signal
analysis and classification through the learning of various
feature space forms [15].
Compared with the traditional fully-connected neural net-
works (FCNNs), CNNs greatly reduce the network structure
parameters, accelerates the training process of the model
and reduces the risk of overfitting to a certain extent [16].
The fundamental advantage of CNNs is that they can be
trained directly using raw temporal signals without the need
for handcrafted feature engineering. In addition, as opposed
to FCNNs, the patterns that CNNs learn are translation-
invariant and they can learn spatial hierarchies of patterns,
which allows them to efficiently learn increasingly complex
and abstract concepts.
deep learning is Rectified Linear Unit (ReLU). The function
returns 0 for negative inputs, but if the input is positive, it
returns that positive value.
The pooling layer takes the feature maps produced by
a convolution layer and extracts the dominant features.
There are two common pooling strategies: max pooling and
average pooling. Max pooling extracts windows from the
input feature maps and returns the highest value whereas
average pooling outputs the average value. In this manner, the
number of learning parameters and the computation burden
are reduced [18].
The flatten layer converts the input data into one-
dimensional feature vector that FC layer can use. In neural
networks, FC layers, also referred to as linear or dense layers,
are frequently used. They link every input neuron to every
output neuron.
III. P ROPOSED SPEED ESTIMATION METHOD
A. Dataset

B. CNN architecture The proposed 1D CNN model will be evaluated on the

The architecture of a CNN consists of four basic elements:
Convolutional layer (CL), Pooling layer (PL), Flatten layer
(FL) and a Fully Connected (FC) layer. The depth of the
model depends on the number of connected layers and
typically the FC layers are the last in that series. Figure 1
depicts the basic blocks of the 1D CNN architecture.
The fundamental operation of CNN is convolution, a linear
operation that involves the multiplication of an input with
convolution filters (a set of weights). A repeated application
of the same filter to the input results in a map of activations
called a feature map, indicating the locations and strength
of a detected feature in an input [17]. The convolutional
layer involves the definition of a variety of hyperparameters,
including stride, padding, type of the activation function,
number and size of filters, etc. The activation function is
very significant since it performs non-linear transformation to
the input enabling the network to improve the representation
ability. One of the most popular activation functions in
VS13 dataset. The dataset contains thirteen vehicles, selected
to be as diverse as possible in terms of manufacturer, produc-
tion year, engine type, power and transmission, resulting in a
total of 400 annotated audio-video recordings [19]. Recorded
speeds in VS13 range from 30 km/h to 110 km/h, as listed
in Table I in [19].
From the original audio-video recordings, 10 second-long
audio sequences in WAV format with a sampling rate of
44 100 Hz were extracted. The speed and camera pass-by
instant of the considered vehicles are recorded in the annota-
tion data. We consider audio segments taken 3 seconds before
and 3 seconds after the pass-by camera instant. Therefore,
the input to the proposed 1D CNN model is one-dimensional
array of length 264 600, as shown in Table I.
B. 1D CNN architecture
The raw sound signal produced by vehicle is used as input
to the proposed 1D CNN model. The architecture of the
model, presented in Table I, is made up of 8 convolutional
layers. The first convolutional layer uses 32 kernels, with
size 32, in order to extract dominant features and suppress
high frequency noise [20]. Afterwards, the kernel size and
the stride size are decreased down the layers. In order to
gain better feature representation, successive small kernels
with size 2 have been used. Average pooling is implemented
right after two consecutive convolutional layers to reduce the
size of the feature space and learn location-invariant signal
characteristics. After flattening the resulting convolutional
output, we stack two fully connected layers, one with 100
and the other with 10 neurons. The output of the network is
one neuron, which gives the speed estimation in kilometer
per hour (km/h). All convolutional layers in the model use
the ReLU activation function, whereas the fully connected
layers have no activation applied.
C. Training procedures
In this paper, we compare two different training ap-
proaches for the proposed 1D CNN model. We used the
Adam optimizer with the learning rate of 0.001. Since the
dataset size is relatively small, we use iterated k-fold cross-
validation (CV) with shuffling, in order to evaluate our model
as precisely as possible and thus ensure better generalization
of the model. We repeat k-fold CV multiple times, for
instance M , shuffling the dataset before splitting it into
K folds. More precisely, we use K = 4, i.e., we repeat
4-fold cross-validation M = 10 times, making a total of
K × M = 40 separate model trainings. Due to shuffling,
in each training, the validation dataset differs from all other
trainings. In both training strategies, in one CV round, one
fold (vehicle) is held out for testing, whereas the remaining
twelve are used for training and validating the model. The
model is implemented in Keras.
In the first approach, we record the training and validation
losses after each training-validation split, in each repetition,
in order to average training and validation losses over all
repetitions. Figure 2 shows the training and validation losses
for Kia Sportage. Similar trend of losses is obtained for
other vehicles in VS13. Due to large errors at the beginning
of the training, we limit y-axis below 200, so that we can
clearly see the trend of the losses. It can be noted that the
training error decreases over the epochs, as expected. On
the other hand, the validation error decreases until it reaches
minimum (optimal value). From that point on, the model
starts to overfit. In Fig. 2, dashed green line denotes the
optimal epoch. We use this epoch to stop the training process
before overfitting and achieve better generalisation on the test
data.
The second training approach focuses on finding a model
with optimal parameters in terms of minimal validation loss.
That is, during each model training (out of 40 trainings), the
validation loss is monitored to ensure saving model weights
that produce minimal validation loss (implemented by the
ModelCheckpoint callback in Keras). Such an obtained
model is compared with the currently best model in terms of
validation loss. If it outperforms the currently best model, it
becomes the currently best model. Otherwise, we neglect it.
Fig. 2. Training and validation losses for Kia Sportage.
This procedure yields a model with minimal validation loss
over 40 trainings with different combination of files used for
validation.
IV. E XPERIMENTAL RESULTS
Vehicle speed estimation will be considered a regression
problem. In regression tasks, the root mean squared error
(RMSE) is typically used to calculate how well the model
fits a dataset. The RMSE can be calculated as
s
1 X
RMSE(y) = (f (xi ) − yi )2 , (1)
N i
where N represent the number of speed measurements, yi is
the vehicle’s actual speed in the i-th measurement and f (xi )
is its corresponding prediction.
The speed estimation performance of the considered train-
ing approaches, given in Section III-C, are compared and
the RMSEs of speed estimations are presented in Table
II, per vehicle and average (bottom). The average speed
estimation errors are 9.65 km/h and 8.88 km/h for the
first (optimal epoch) and the second approach (model with
minimal validation loss), respectively.
It can be seen that Renault Scenic performs significantly
worse than the other vehicles in terms of speed estimation
for both approaches. This is expected since Renault Scenic
recording session took place on a windy day, which resulted
in a strong wind corrupting the sound of the vehicle [19]. On
the other hand, Peugeot 208 has the lowest speed estimation
error for both cases.
Compared with [11], which yields RMSE = 7.39 km/h,
obtained on a subset of 10 out of 13 vehicles from VS13,
our second approach gives notably higher RMSE. On the
other hand, the advantage of our approach is that is does not
require any human-crafted feature, such as the MA feature
introduced in [11]. The 1D CNN model proposed herein can
be trained end-to-end, which is a considerable advantage with
respect to [11]. The obtained results are promising and set a
baseline for future research.
V. C ONCLUSIONS
We proposed an end-to-end 1D CNN model for vehicle
speed estimation which uses raw sound signal, and gives
an average estimation error of 8.88 km/h on the VS13
 TABLE II
[5] A. Yabo, S. I. Arroyo, F. G. Safar, and D. Oliva, “Vehicle classification
RMS S OF SPEED ESTIMATION and speed estimation using computer vision techniques,” in XXV
Congreso Argentino de Control Automático (AADECA 2016)(Buenos
VEHICLE RMSE1 [km/h] RMSE2 [km/h] Aires, 2016), 2016.
Citroen C4 Picasso 7.83 7.15 [6] M. Yan, M. Li, H. He, and J. Peng, “Deep learning for vehicle speed
Kia Sportage 11.38 7.93 prediction,” Energy Procedia, vol. 152, pp. 618–623, 2018.
Mazda 3 9.22 8.48 [7] H. V. Koops and F. Franchetti, “An ensemble technique for estimating
Mercedes AMG550 10.44 8.5 vehicle speed and gear position from acoustic data,” in 2015 IEEE
Mercedes GLA 11.32 8.24 International Conference on Digital Signal Processing (DSP). IEEE,
Nissan Qashqai 10.17 5.53 2015, pp. 422–426.
Opel Insignia 8.00 9.75 [8] S. Barnwal, R. Barnwal, R. Hegde, R. Singh, and B. Raj, “Doppler
Peugeot 208 5.81 4.71 based speed estimation of vehicles using passive sensor,” in 2013
Peugeot 3008 10.70 10.25 IEEE International Conference on Multimedia and Expo Workshops
Peugeot 307 9.49 8.09 (ICMEW). IEEE, 2013, pp. 1–4.
Renault Captur 7.94 8.23 [9] E. Kubera, A. Wieczorkowska, A. Kuranc, and T. Słowik, “Discov-
Renault Scenic 15.38 19.96 ering speed changes of vehicles from audio data,” Sensors, vol. 19,
no. 14, p. 3067, 2019.
VW Passat 7.85 8.62
[10] J. Giraldo-Guzmán, A. G. Marrugo, and S. H. Contreras-Ortiz, “Ve-
Average 9.65 8.88 hicle speed estimation using audio features and neural networks,” in
2016 IEEE ANDESCON. IEEE, 2016, pp. 1–4.
[11] S. Djukanović, J. Matas, and T. Virtanen, “Acoustic vehicle speed
estimation from single sensor measurements,” IEEE Sensors Journal,
dataset. The main limitation of the proposed method for vol. 21, no. 20, pp. 23 317–23 324, 2021.
sound speed estimation is the size of the training dataset. [12] N. Bulatović and S. Djukanović, “An approach to improving sound-
based vehicle speed estimation,” in 2022 Zooming innovation in
In the context of deep learning, a set of 400 samples is consumer technologies conference (ZINC 2022). IEEE, 2022.
small to achieve reliable results. Cross-validation techniques [13] N. Bulatović and S. Djukanović, “Mel-spectrogram features for acous-
can help to a certain extent to overcome this limitation, tic vehicle detection and speed estimation,” in 2022 26th International
Conference on Information Technology (IT). IEEE, 2022, pp. 1–4.
however a sufficient dataset increase is still necessary to [14] S. Kiranyaz, O. Avci, O. Abdeljaber, T. Ince, M. Gabbouj, and D. J.
make substantial progress. In this sense, one of the future Inman, “1d convolutional neural networks and applications: A survey,”
research directions will be extension of the dataset with data Mechanical systems and signal processing, vol. 151, p. 107398, 2021.
[15] LJ. Stanković and D. Mandić, “Convolutional neural networks demys-
augmentation techniques. We will also consider increasing tified: A matched filtering perspective-based tutorial,” IEEE Transac-
the dataset by adding samples from another dataset, used in tions on Systems, Man, and Cybernetics: Systems, 2023.
earlier research on the topic of vehicle counting [21], [22]. [16] C. Du, R. Zhong, Y. Zhuo, X. Zhang, F. Yu, F. Li, Y. Rong, and
Y. Gong, “Research on fault diagnosis of automobile engines based
Those samples could be annotated using vision-based speed on the deep learning 1d-CNN method,” Engineering Research Express,
estimation methods. vol. 4, no. 1, p. 015003, 2022.
[17] J. Brownlee, “How do convolutional layers work in deep learning
R EFERENCES neural networks,” Machine Learning Mastery, vol. 17, 2020.
[18] S. Sharma and R. Mehra, “Implications of pooling strategies in convo-
[1] G. Radivojevć, G. Šormaz, and B. Lazić, “Trends in development lutional neural networks: A deep insight,” Foundations of Computing
and integration of ITS,” Journal of Road and Traffic Engineering, and Decision Sciences, vol. 44, no. 3, pp. 303–330, 2019.
vol. 66, no. 3, pp. 33–40, Oct. 2020. [Online]. Available: [19] S. Djukanović, N. Bulatović, and I. Čavor, “A dataset for audio-video
https://www.putisaobracaj.rs/index.php/PiS/article/view/151 based vehicle speed estimation,” in 2022 30th Telecommunications
[2] C. Wilson, C. Willis, J. K. Hendrikz, R. Le Brocque, and N. Bellamy, Forum (TELFOR). IEEE, 2022, pp. 1–4.
“Speed cameras for the prevention of road traffic injuries and deaths,” [20] W. Zhang, G. Peng, C. Li, Y. Chen, and Z. Zhang, “A new deep
Cochrane database of systematic reviews, no. 11, 2010. learning model for fault diagnosis with good anti-noise and domain
[3] J. Kurniawan, S. G. Syahra, C. K. Dewa et al., “Traffic congestion adaptation ability on raw vibration signals,” Sensors, vol. 17, no. 2, p.
detection: learning from CCTV monitoring images using convolutional 425, 2017.
neural network,” Procedia computer science, vol. 144, pp. 291–297, [21] S. Djukanović, J. Matas, and T. Virtanen, “Robust audio-based vehicle
2018. counting in low-to-moderate traffic flow,” in 2020 IEEE Intelligent
[4] H. Dong, M. Wen, and Z. Yang, “Vehicle speed estimation based on Vehicles Symposium (IV). IEEE, 2020, pp. 1608–1614.
3d convnets and non-local blocks,” Future Internet, vol. 11, no. 6, p. [22] S. Djukanović, Y. Patel, J. Matas, and T. Virtanen, “Neural network-
123, 2019. based acoustic vehicle counting,” in 29th European Signal Processing
Conference (EUSIPCO 2021), 2021.