Article citation information:
Pamuła, W.,
Kłos, M.J. On
site processing of video stream for mapping traffic parameters. Scientific Journal of Silesian University of
Technology. Series Transport. 2022, 117,
175-189. ISSN: 0209-3324. DOI: https://doi.org/10.20858/sjsutst.2022.117.12.
Wieslaw PAMULA[1], Marcin Jacek KŁOS[2]
ON SITE PROCESSING OF VIDEO STREAM FOR MAPPING TRAFFIC PARAMETERS
Summary. Traffic
surveillance provides crucial data for the operation of intelligent
transportation systems. The growing number of cameras in the transport system
poses a problem for the efficient processing of surveillance data. Processing
of video data for extracting traffic parameters is usually done using image
processing methods and requires substantial processing resources. An
alternative way is to transform the video stream and map the traffic parameters
using the obtained transform coefficients. Spatiotemporal wavelet transform of
the video stream contents, using filter banks, is proposed for mapping traffic
parameters. Performed tests prove good resilience to illumination changes of
road scenes. Mapping errors are smaller than in the case of the commonly used
video detectors at sites on multilane roads with low to moderate traffic load.
Keywords: video
surveillance, discrete wavelet transforms, traffic flow, traffic density,
intelligent transportation systems
1. INTRODUCTION
Intelligent
transportation systems integrate information and communication technologies to
improve the functioning of road networks and increase the efficiency of moving
people and goods [1].
Determining the state of the transport system is decisive for the development
of traffic control decisions [2].
Modelling of state changes contributes to establishing traffic management
strategies [3,4].
Traditional approaches to measuring road traffic parameters, which constitute
the bulk of information on the transport systems state, incorporate inductive
loop detectors, magnetic detectors, ultrasonic detectors, radar detectors, and
laser detectors [5,6].
Most of these devices can provide only spot data, which means data collected at
defined areas of traffic lanes.
The
development of intelligent transportation systems is manifested by the growing
number of cameras distributed on the road network [7,8].
Hence, it is essential to integrate them into a functional entity to provide
traffic data. The large volume of raw image data can pose an acute problem for
the efficient processing of the content. Video cameras provide rich contextual
information on the course of road traffic. Image processing-based methods are
mostly used for extracting traffic parameters [9–11].
The accuracy and reliability of measurements, in this case, are coincident with
the complexity of applied image processing algorithms and necessary computing
resources. Detection of individual vehicles is a preliminary stage of
determining road traffic parameters. The variable parameters of the
observation, especially changes in illumination, pose problems for correct
detection [12,13].
Background modelling is used to determine the empty road template, which is
subtracted from the current video frames. The result depicts
moving objects, by default, vehicles [14,15].
Several models developed for different observation environments require modest
processing power for implementation and achieve proper vehicle detection mainly
for limited changes of the road view [16,17].
The
extraction of low-level features of images increases the processing complexity
for detecting objects. This approach involves the application of filters and
the clustering of filter results. Detection results surpass background
subtraction methods but still fall short of expectations in the case of highly
changing illumination of the traffic scene. Local feature descriptors, such as
Histograms of Gradients (HOG)[11],
Scale Invariant Feature Transform (SIFT) [18],
Speed-Up Robust Feature (SURF), and Gradient Location Orientation Histogram (GLOH) [19]
improve the detection capabilities but impose still higher processing
requirements [20,21].
Vehicle
motion models once again introduce a higher requirement on processing power.
The models are based on the calculation of optical flow and connected region
analysis. Horn–Schunck optical flow estimation
algorithm [22] is
the starting point of modifications in the course of increasing the robustness
of vehicle detection. For instance, Peng et al. proposed the use of inter-frame
differences for triggering calculations, which significantly reduces the
computation burden for updating the optical flow field [23].
Wavelet-based
transforms are readily used for the analysis of road traffic parameters. The
ability of localizing features and multi-resolution representation of parameter
changes are arguments for the application of these transforms. The input data
for analysis are predominantly a series of measurements collected at sites of
the road network. These are time instants of detections of individual vehicles,
vehicle speed values, distances between vehicles, or aggregated quantities. The
registered data points are indexed in time order and can be regarded as time
series data. Time series analysis comprises frequency domain methods and time
domain methods. Wavelet analysis belongs to the frequency domain group of
methods. Two main problems are studied: designation of the wavelet basis
function of the transform and determination of the level of decomposition, for
effective description of the traffic data. The “ability” of
a wavelet to represent different features of traffic data is characterised
by the size of the support and the number of vanishing moments, whereas the
decomposition level delimits the resolution for the extraction of data
attributes.
Early
works [7,24]
concentrated their efforts on the analysis of traffic flow patterns. The
statistical autocorrelation function (ACF) is used
for the selection of the decomposition level. This function is usually used to
detect trends and seasonality in a time series. In this case, ACF is calculated for the original dataset and wavelet
decompositions at different levels of the dataset. Equal ACF
values signify the correct choice of the decomposition level. In [25], was
proposed the wavelet transform of loop detector data for revealing bottlenecks,
transient traffic, and traffic oscillations. Wavelet-based energy peaks from
vehicle to vehicle are traced. The duration and intensity of the peaks are
processed to obtain traffic features and calculate traffic parameters [25,26].
The task of detecting singularities in noisy traffic data is studied,
singularities in traffic data may indicate bottlenecks or traffic incidents.
The
problem of video-based traffic surveillance is addressed in [27].
The authors use a two-dimensional discrete wavelet transform for extracting
features describing vehicles from the images. Haar
wavelet is used as the basis, and the decomposition is done in the space
domain. Tests using highway traffic images prove good resilience to shadows on
the traffic lanes.
In
this paper, spatiotemporal wavelet transform of the video stream from the
observation camera is proposed for mapping road traffic parameters instead of
applying image processing. The change of contents of the stream represented by
transform coefficients, instead of detecting and tracking vehicles, is the
basis of the mapping. The mapping of video for representation to traffic
parameters is not reported. Reported methods for crowd analysis using video
from surveillance cameras share components of this approach. Crowd density is
determined using direct processing of video content [28].
The
digital form of video data imposes the use of discrete wavelet transform (DWT)
versions of the transform. Preliminary tests show that the use of wavelet-based
transform of video data retains the characteristics of traffic parameters. A
set of detection fields is defined on the image of an observed traffic lane.
Passing vehicles are recorded entering these areas. The weighed sum of
coefficients of the wavelet transform of a vehicle detection field content
corresponds to the traffic density observed on the traffic lane; a similar
correspondence is observed for traffic flow.
The
primary objective of this paper is to present the idea of the method for the
application of the spatiotemporal wavelet transform to map road traffic
parameters such as traffic flow and traffic density.
2.
MAPPING ROAD TRAFFIC PARAMETERS
The
problem of mapping road traffic parameters using transform coefficients of a
video stream is the goal of this study. What parameters of a wavelet transform
give a good estimation of the road traffic parameters such as traffic flow and
traffic density?
The
domain of wavelet transforms is chosen as the basis for this study. The
literature review gives examples of feature extraction, especially space
features, for finding relations with traffic parameters. The proposed idea
focuses on the temporal features of the video stream. Decomposition of the
video stream in time enables the extraction of features that can describe road
traffic. Additional space decomposition reduces the data stream for processing.
The task is to determine the levels of decomposition and choose the
coefficients that are significant for mapping road traffic. The mapping results
should not diverge substantially from the mappings obtained by commonly used
video-based devices - video detectors. The video stream to be processed comes
from surveillance cameras mounted above roads leading to town centres. No
special cameras were used for the video data collection.
2.1.
Road scene model
Traffic
flow and traffic density values carry the most important information useful for
controlling and managing road traffic. These parameters are the objects of
mapping. The input to the mapping is a video stream depicting the changing
traffic scene. CCTV cameras are the usual source of video data. Important
parameters of the stream are resolution, range of observation, and speed of
registration. Video stream
The
values of
Road
traffic combines the movement of vehicles of different sizes and with various
dynamic properties. To capture these characteristics, a multi-resolution
representation is proposed as the basis for mapping traffic parameters. This
approach is related to finding description keys at distinct scales of
observation of the traffic. Such descriptions can be nested [29].
Techniques to compute nested sequences of multi-resolution representations are
closely related to wavelets. Multiscale representation using wavelets was
introduced by S. Mallat in [30].
2.2. Description
of video stream contents
Description
of image contents is done using a two-dimensional spatial discrete wavelet
transforms. To capture changes of the video stream in time, the transforms are
extended to include processing in time. The video stream is represented using
wavelet coefficients
|
|
(2) |
The
dyadic scale is used, the scaling function
|
|
(3) |
Where
j – scale, k, m, n – shifts, all integers, the mother wavelet is
shifted and scaled by powers of 2.
Separable
wavelet functions are used for transforming the video stream, in this case, the
functions can be rewritten, exposing the 1D
components. There are seven combinations
of φ and ψ:
|
|
(4) |
Efficient
computing of coefficients is carried out using filters. Transform functions are
substituted by filters defined by sets of weights corresponding to the
characteristics of the functions. Filter
·
approximation
|
|
(5) |
·
details
|
|
(6) |
where
The
approximation coefficients are decomposed with combinations of filters and then
down sampled. This is represented as a filter bank in Figure 1.
Tab.
1
Prediction and update
functions for Deslauriers-Dubuc wavelets
|
Wavelet |
Prediction and update functions |
|
DD(1,1) |
|
|
|
|
|
DD(2,2) |
|
|
|
|
|
DD(4,4) |
|
|
|
Table
1 lists the wavelet transforms used in this study. The least demanding
computationally wavelet DD(1,1) corresponds to the Haar wavelet. Deslauriers-Dubuc
interpolating scaling functions, also known as Interpolets,
are good candidates for such applications. To streamline calculations, the
lifting scheme is used. The lifting step consists of
|
|
(7) |
The
choice of wavelet basis functions for the transforms is conditioned by the
complexity of the calculation. Effective solutions, for instance, incorporating
logic-based processing, suitable for on-site designs [9],
call for integer based calculations.
2.3. Method
for mapping of road traffic parameters
The
calculation of spatiotemporal wavelet transform coefficients according to Mallat's scheme results in a set of detail
coefficients for every decomposition level and one set of approximation
coefficients for the last decomposition level. Decomposition level labels the
consecutive step of application of the set of filters to
The
lower the decomposition level, the larger the number of coefficients in the
set. The question arises, which coefficients carry significant clues for
describing traffic parameters and which can be discarded. If the length of the
transformed stream is
Only
the coefficients at the highest level of decomposition are used for mapping.
These carry the synthetic description of the image contents changes at patches
of the size (512/23) and in a period of
|
|
(8) |
Weight
values
The
observed traffic lane is represented by sums of coefficients
|
|
(9) |
These
are calculated for a number of periods
|
|
(10) |
For
|
|
(11) |
which gives:
|
|
(12) |
The
derived set of weights is specific for a measurement site.
3. RESULTS AND
COMPARISON WITH PERFORMANCE OF VIDEO DETECTORS
Mapping
of traffic flow and traffic density using wavelet transform coefficients is
examined in this section. Road traffic data collected at several sites are used
for calculating the weights of the representations (equation 12). Figure 1
shows examples of camera sites where multilane roads with high to low traffic
loads are observed.
Fig.
1. Camera sites views: a) high traffic, b) medium traffic, c) low traffic
The
range of observation is limited by the acceptable sizes of vehicles expressed
in the pixels of the image. Image resolution and noise level impose the
condition that the smallest vehicle size should be a few hundred pixels. This
defines a field of view not longer than 150 metres when a standard CCTV camera
is used. The highest level of wavelet transform decomposition in the space of
the image is determined by the need to preserve vehicle representations. A CCTV
image of the size 720
The
level of wavelet transform decomposition in time
3.1. Measurement
sites
Data
from three measuring sites localized on multilane roads are used for the tests
– Figure 1. The sites have highly illuminated and shadowed lanes.
Vehicles travelling on the illuminated road lane cast shadows on the parallel
lane, causing it to be shadowed. The parallel lane contains combined shadows
from both traffic lanes, and this is a source of errors. The moving sun may
change the proportion of shadows on the lanes if the lanes are north- or
south-bound; in this case, the proportions are constant. The lane with more
shadows is named shadowed.
Fig.
2. Graphs of traffic flow values recorded at the measurement sites
This
shadowing phenomenon is of particular interest as the source of measurement
errors. In the case of video detectors, it usually causes extra vehicle
detections. Morning traffic parameters are measured. Three measurement sites
differing in the size of the traffic flow are selected: high, medium, and low
traffic flow. Figure 2 illustrates the changes in values of traffic flow at the
measurement sites during the measurement time period. High traffic flow
surpasses 1800 veh./h per
lane, whereas the site with low traffic has flow values below 500 veh./h per lane. The largest flow differences between these
two road lanes are noted at site (b) with the medium volume of traffic.
3.2. Comparison of results of the
proposed method
The
proposed approach for mapping traffic parameters is compared with the
performance of video-based measurement devices present at traffic sites. Two
devices are chosen to represent the current state of video-based vehicle
detection technology. Both devices use basic image processing techniques for
determining the presence of objects in the detection fields predefined on
observed images of the road. The objects assumed to be vehicles are counted,
and the times of their entry and duration of presence in the detection fields
are used to calculate traffic parameters such as flow and density.
The
algorithm of detection of the first device, A, tracks the content of the detection
field and when it substantially differs from the model of the observed
background an object’s presence is signalled. The background is modelled
statistically using one probability distribution. An example of the
implementation of this principle of operation is protected by a patent [29].
The second device – B uses a more complex detection principle in which
both the background and the detection fields are modelled using a fuzzy-based
feature update algorithm, and when the two models differ an object’s
presence is signalled.
The
carried out mappings using DD(2,2) and DD(4,4) show no
significant performance advantage over DD(1,1). The calculation of wavelet
coefficients requires, in these instances, much more processing resources,
which impairs on-site implementations. Representative results of mapping
traffic parameters using DD(1,1) are further
discussed. Traffic flow and traffic density are mapped using coefficients of DD(1,1) transforms of video streams of road scenes.
3.3. Traffic flow
DD(1,1) with varying decomposition
parameters is used for mapping traffic flow. Second, third and fourth space
decomposition levels and 13th and 14th temporal decomposition levels are
investigated. The best mapping results are obtained for the set (14,3,3). None of the transform coefficients explicitly
outweighs the others; this indicates that the camera observation parameters
decisively determine the weights. Larger values of weights are noted for
mapping most of the flow values on highly illuminated road lanes than on corresponding
shadowed lanes.
The
video database consists of non-compressed films of road lanes recorded at the
measurement sites. This material is inputted to the vehicle detection devices
– video detectors, in real-time, and the detection results are recorded. Standard
detection settings were used. The obtained values are matched with the
reference sets of traffic parameter values.
Table
2 summarizes the results of measuring traffic flow. MAPE
error values for all measurement sites are lower than RMSPE
values. MAPE values are less sensitive to outliers in
comparison to RMSE values. The difference does not
exceed ¼ of the MAPE, indicating a few
outliers.
Tab. 2
Values
of mapping errors for traffic flow
|
Mapping errors [%] |
Measurement sites |
||||||
|
High traffic |
Medium traffic |
Low traffic |
|||||
|
Light |
Shadow |
Light |
Shadow |
Light |
Shadow |
||
|
Video detector A |
RMSPE |
25 |
26 |
27 |
13 |
15 |
16 |
|
MAPE |
24 |
22 |
24 |
10 |
14 |
15 |
|
|
Video detector B |
RMSPE |
10 |
4,9 |
19 |
6,7 |
19 |
35 |
|
MAPE |
8,9 |
3,8 |
14 |
5,0 |
16 |
33 |
|
|
Proposed method |
RMSPE |
5,9 |
7,8 |
21 |
16 |
9,9 |
20 |
|
MAPE |
3,8 |
4,1 |
16 |
12 |
8,0 |
15 |
|
Traffic
on illuminated lanes is more accurately mapped than on the shadowed lanes,
except in the case of medium traffic. The graph in Figure 3 shows that at the
medium traffic site, traffic flow changes are more volatile than at the other
sites. Examination of the video shows that large errors arise when vehicles
temporarily slow down or stop due to abrupt changes in traffic density (traffic
jams), and this is not captured by the transform. Some errors are caused by
container trucks travelling in bunches. Higher placement of the observation
camera can remedy this weakness.
Fig.
3. Mapping errors of traffic flow values
Differences
in RMSPE and MAPE error
values are small, although video detector B shows a larger number of outliers.
Video detector A copes better with low traffic, while
video detector B with high traffic. The proposed transform-based processing
performs better than the video detectors. There are no outstanding error
values. Box plots presented in Figure 3 illustrate the error statistics in
detail.
Video
detector B gives smaller errors than the mappings, but there are numerous
outliers. Detailed inspection of the video material shows that these are the
consequences of stopped vehicles, as it is in the case of mappings but the
results generate much higher error values.
3.4. Traffic density
Again
wavelet DD(1,1) with varying decomposition parameters
is used for mapping traffic density. The same range of decomposition parameters
is applied. The best mapping results are obtained for the set (14,3,3). Table 3 presents the errors in mapping traffic
density. In comparison to flow mapping weights, the density mapping weights are
substantially different. Some weights have very small values for all examined
measurement sites. This can be of use in optimizing processing operations for
calculating traffic density.
Tab. 3
Values
of mapping errors for traffic density
|
Mapping errors [%] |
Measurement sites |
||||||
|
a) High traffic |
b) Medium traffic |
c) Low traffic |
|||||
|
Light |
Shadow |
Light |
Shadow |
Light |
Shadow |
||
|
Video detector A |
RMSPE |
10 |
14 |
9,5 |
9,5 |
17 |
38 |
|
MAPE |
7,7 |
12 |
7,0 |
7,9 |
14 |
32 |
|
|
Video detector B |
RMSPE |
11 |
6,0 |
22 |
8,1 |
17 |
38 |
|
MAPE |
10 |
5,2 |
17 |
6,6 |
12 |
26 |
|
|
Proposed method |
RMSPE |
9,0 |
19 |
14 |
14 |
12 |
22 |
|
MAPE |
7,2 |
15 |
12 |
11 |
10 |
15 |
|
Mapping
errors follow the same pattern as in the case of traffic flow. In all, errors
are larger, especially at the high traffic site. Box plots presented in Figure
4 illustrate the error statistics in detail.
Fig.
4. Mapping errors of traffic density values
Differences
in RMSPE and MAPE error
values are small for the proposed method. Traffic density at low traffic sites
on shadowed lanes is determined with the largest errors by both video
detectors. This poor performance may be linked to losing infrequently passing
vehicles due to inadequate detection ability of objects partially covered by
the shadows, which disrupt the object’s view.
4. DISCUSSION
Table
4 summarises the comparison of the performance of video detectors and wavelet
mappings. The advantage of the proposed method is not significant but the
consistency of the mapping - there are no outliers, is important for traffic
control and management systems.
The
processing algorithm of video detector B presumably loses vehicles due to poor
sensitivity to infrequently passing objects on the image. This may be caused by
the parameters of updating the background model in the device. Similarly,
several outliers in the case of medium traffic at an illuminated site also
suggest that such conditions pose momentary difficulties in discerning and
tracking features.
Tab. 4
Average errors in mapping and measuring traffic parameters for all sites
|
Average errors [%] |
Video detector A |
Video detector B |
Proposed method |
|||
|
flow |
density |
flow |
density |
flow |
density |
|
|
RMSPE |
20 |
16 |
16 |
17 |
13 |
15 |
|
MAPE |
18 |
13 |
13 |
13 |
10 |
12 |
The
proposed method maps traffic flow and traffic density more accurately than
commonly used video-based vehicle detection devices. In the case of high and
low traffic, the ranges of errors are substantially lower. High momentary
errors are recorded when the video contents reveal a stopped vehicle, which
caused numerous lane changes by vehicles approaching this obstacle. High errors
are also caused by queues of container trucks. These situations are less
effectively represented by the transform coefficients especially related to
temporal changes. The coefficient values indicate the scale of changes in time
at different time resolutions. A high level of decomposition diminishes
the sensitivity to high speed changes of contents, which are induced by such
traffic situations.
Large
vehicles present in the traffic lanes cause error fluctuations. Another level
of decomposition can be chosen to alleviate the deficiency of different size
object mapping in the course of transforming the video data. This approach
should consider the characteristics of the observed road, that is, whether it
is a transit road with heavy vehicles or an urban road mainly with cars.
The
errors in measuring traffic density are higher than traffic flow; it can be
attributed to the higher influence of illumination changes in deriving the
results. Modelling background as well as calculating wavelet coefficients is
susceptible to noise. Changing illumination values can be regarded as a noise
factor with highly volatile probability distribution parameters.
The
advantage of the transform-based approach lies in the reduction of computing
operations for obtaining the mapping of traffic parameters. For instance,
background subtraction requires background modelling involving statistical
calculations using image pixel neighbourhoods that are hundreds of calculations
per image pixel. Observation cameras provide video streams with a resolution of
720
Transform
calculations may be done in a processing pipeline using a non-processor based
device. Implementation of calculations in all requires tens of operations per
pixel, which are performed in parallel, at the speed of the incoming pixels.
5. CONCLUSIONS
The
proposed method enables the mapping of road traffic parameters on multilane
roads with smaller errors than the solutions currently implemented in video
detecting devices. The video detecting devices perform poorly, especially when
the road image is corrupted by shadows of vehicles travelling on adjacent
traffic lanes.
The
spatiotemporal wavelet transform, by selecting different decomposition
parameters, allows for the representation of features at different resolutions
in time and space. It represents the features of objects at different scales -
by choosing a decomposition level, it is possible to "filter out" vehicles
with different sizes or characteristic details of appearance. This makes it
possible to identify the position of individual vehicles in the video stream.
The time transformation describes the dynamics of changes in the movement of
the vehicles. This information is useful for mapping the changes in traffic
parameters.
The
discrete wavelet transform can be implemented using the lifting scheme,
significantly reducing the required computation budget. The application of an
embedded processing system comprising of a field programmable gate array can
efficiently calculate transform coefficients in real-time. Such one chip
solutions can be integrated with traffic monitoring cameras and function as
traffic data collection subsystem nodes in intelligent transportation systems.
Finally,
the proposed method of mapping road traffic parameters proves that a set of
weighed coefficients of a wavelet transform give a credible estimation of road
traffic parameters, such as traffic flow and traffic density. Hence, the
proposed method requires further studies in the optimization of the processing
algorithms suitable for available hardware resources.
References
1. Bok
Jinjoo, Youngsang Kwon. 2016. „Comparable Measures of Accessibility to Public Transport Using the General
Transit Feed Specification.” Sustainability 8(3). DOI: 10.3390/su8030224.
2. Rith
Monorom, Alexis Fillone, Jose Bienvenido M. Biona. 2019. „The Impact of Socioeconomic Characteristics and Land
Use Patterns on Household Vehicle Ownership and Energy Consumption in an Urban
Area with Insufficient Public Transport Service – A Case Study of Metro
Manila.” Journal of Transport Geography 79: 102484. DOI: 10.1016/j.jtrangeo.2019.102484.
3. Zhang
Tingru, Alan H.S. Chan, Hongjun Xue, Xiaoyan Zhang, Da Tao. 2019.
“Driving Anger, Aberrant Driving Behaviors, and Road Crash Risk: Testing
of a Mediated Model.”International Journal of Environmental Research
and Public Health 16(3): 1-13. DOI: 10.3390/ijerph16030297.
4. Ortega
Jairo, János Tóth, Tamás Péter. 2021.
„Planning a Park and Ride System: A Literature Review.”Future
Transportation 1(1): 82-98. DOI: 10.3390/futuretransp1010006.
5. Federal
Highway Administration. 2016. „Traffic Monitoring Guide FHWA.”
Fhwa. Available at: http://www.fhwa.dot.gov/policyinformation/tmguide/.
6. Klein
Lawrence A. 2017. ITS Sensors and
Architectures for Traffic Management and Connected Vehicles. Boca
Raton : Taylor & Francis, CRC Press. DOI: 10.1201/9781315206905.
7. Jiang
Xiaomo, Hojjat Adeli. 2004. „Wavelet Packet-Autocorrelation Function
Method for Traffic Flow Pattern Analysis.” Computer-Aided Civil and Infrastructure Engineering 19(5): 324-37. DOI: 10.1111/j.1467-8667.2004.00360.x.
8. Mandellos
Nicholas A., Iphigenia Keramitsoglou, Chris T. Kiranoudis. 2011. „A Background Subtraction Algorithm for Detecting
and Tracking Vehicles.” Expert Systems with Applications 38(3):
1619-31. DOI: 10.1016/j.eswa.2010.07.083.
9. Tasgaonkar
Pankaj P., Rahul Dev Garg, Pradeep Kumar Garg. 2020. „Vehicle Detection
and Traffic Estimation with Sensors Technologies for Intelligent Transportation
Systems.” Sensing and Imaging 21(1). DOI: 10.1007/s11220-020-00295-2.
10. Singleton
Patrick A., Keunhyun Park, Doo Hong Lee. 2021. „Varying Influences of the
Built Environment on Daily and Hourly Pedestrian Crossing Volumes at Signalized
Intersections Estimated from Traffic Signal Controller Event Data.” Journal
of Transport Geography 93: 103067. DOI: 10.1016/j.jtrangeo.2021.103067.
11. Buch Norbert,
Sergio A. Velastin, James Orwell. 2011. „A Review of Computer Vision
Techniques for the Analysis of Urban Traffic.” IEEE Transactions on
Intelligent Transportation Systems 12(3): 920-39. DOI: 10.1109/TITS.2011.2119372.
12. Semertzidis
T., K. Dimitropoulos, A. Koutsia, N. Grammalidis. 2010. „Video Sensor
Network for Real-Time Traffic Monitoring and Surveillance.” IET
Intelligent Transport Systems 4(2): 103-12. DOI: 10.1049/iet-its.2008.0092.
13. Roy Arunesh,
Nicholas Gale, Lang Hong. 2011. „Automated Traffic Surveillance Using
Fusion of Doppler Radar and Video Information.” Mathematical and
Computer Modelling 54(1-2): 531-43. DOI: 10.1016/j.mcm.2011.02.043.
14. Xu Yong,
Jixiang Dong, Bob Zhang, Daoyun Xu. 2016. „Background Modeling Methods in
Video Analysis: A Review and Comparative Evaluation.” CAAI
Transactions on Intelligence Technology 1(1): 43-60. DOI: 10.1016/j.trit.2016.03.005.
15. Garcia-Garcia
Belmar, Thierry Bouwmans, Alberto Jorge Rosales Silva. 2020. „Background
Subtraction in Real Applications: Challenges, Current Models and Future
Directions.” Computer Science Review 35: 100204. DOI: 10.1016/j.cosrev.2019.100204.
16. Mitrović
Dejan. 2005. „Reliable Method for Driving Events Recognition.” IEEE
Transactions on Intelligent Transportation Systems 6(2): 198-205. DOI: 10.1109/TITS.2005.848367.
17. Lee Uichin,
Mario Gerla. 2010. „A Survey of Urban Vehicular Sensing Platforms.”
Computer Networks 54(4): 527-44. DOI: 10.1016/j.comnet.2009.07.011.
18. Salih Yasir,
Aamir Saeed Malik. 2011. „Comparison of Stochastic Filtering Methods for
3D Tracking.” Pattern Recognition 44(10-11): 2711-37. DOI: 10.1016/j.patcog.2011.03.027.
19. Karasulu
Bahadir, Serdar Korukoglu. 2012. „Moving Object Detection and Tracking by
Using Annealed Background Subtraction Method in Videos: Performance
Optimization.” Expert Systems with Applications 39(1): 33-43. DOI:
10.1016/j.eswa.2011.06.040.
20. Guo Yulan,
Mohammed Bennamoun, Ferdous Sohel, Min Lu, Jianwei Wan, Ngai Ming Kwok. 2016.
„A Comprehensive Performance Evaluation of 3D Local Feature
Descriptors.” International Journal of Computer Vision 116(1): 66-89.
DOI: 10.1007/s11263-015-0824-y.
21. Kihl Olivier,
David Picard, Philippe-Henri Gosselin. 2015. „A Unified Framework for Local
Visual Descriptors Evaluation.” Pattern Recognition 48(4): 1174-84.
DOI: 10.1016/j.patcog.2014.11.013.
22. Horn Berthold
K.P., Brian G. Schunck. 1981. „Determining Optical Flow.” Artificial
Intelligence 17(1-3): 185-203. DOI: 10.1016/0004-3702(81)90024-2.
23. Peng Yanan,
Zhenxue Chen, Q.M. Jonathan Wu, Chengyun Liu. 2018. „Traffic Flow
Detection and Statistics via Improved Optical Flow and Connected Region
Analysis.” Signal, Image and Video Processing 12(1): 99-105. DOI: 10.1007/s11760-017-1135-2.
24. Adeli Hojjat,
Samanwoy Ghosh-Dastidar. 2004. „Mesoscopic-Wavelet Freeway Work Zone Flow
and Congestion Feature Extraction Model.” Journal of Transportation
Engineering 130(1): 94-103. DOI: 10.1061/(ASCE)0733-947X(2004)130:1(94).
25. Zheng Zuduo,
Soyoung Ahn, Danjue Chen, Jorge Laval. 2011. „Applications of Wavelet
Transform for Analysis of Freeway Traffic: Bottlenecks, Transient Traffic, and
Traffic Oscillations.” Transportation Research Part B: Methodological
45(2): 372-84. DOI: 10.1016/j.trb.2010.08.002.
26. Zheng Zuduo,
Soyoung Ahn, Danjue Chen, and Jorge Laval. 2011. „Freeway Traffic
Oscillations: Microscopic Analysis of Formations and Propagations Using Wavelet
Transform.” Procedia - Social and Behavioral Sciences 17: 702-16. DOI: 10.1016/j.sbspro.2011.04.540.
27. Ibtissam
Slimani, Abdelmoghit Zaarane, Abdellatif Hamdoun, Issam Atouf. 2018.
„Traffic Surveillance System For Vehicle Detection Using Discrete Wavelet
Transfor.” Journal of Theoretical
and Applied Information Technology 96(17). DOI: 10.13140/RG.2.2.34426.08649.
28. Bendali-Braham
Mounir, Jonathan Weber, Germain Forestier, Lhassane Idoumghar, Pierre-Alain
Muller. 2021. „Recent Trends in Crowd Analysis: A Review.” Machine
Learning with Applications 4: 100023. DOI: 10.1016/j.mlwa.2021.100023.
29. US4847772A. Michalopoulos
Panos G., A. Richard Fundakowski, Meletios Geokezas, Robert C. Fitch. 1989. "Vehicle
detection through image processing for traffic surveillance and control". Patent
number 4,847,772. Available at: https://patents.google.com/patent/US4847772A/en.
30. Mallat
Stephane G. 2009. „Multiresolution Approximations and Wavelet Orthonormal
Bases of L2(R).” Fundamental Papers in Wavelet Theory 315(1): 524-42.
DOI: 10.1090/s0002-9947-1989-1008470-5.
Received 02.06.2022; accepted in
revised form 08.09.2022
Scientific Journal of Silesian University of Technology. Series
Transport is licensed under a Creative Commons Attribution 4.0
International License
[1]Faculty of Transportand
Aviation Engineering, The Silesian University of
Technology, Krasińskiego 8 Street, 40-019
Katowice, Poland. Email:
wieslaw.pamula@polsl.pl. ORCID:
https://orcid.org/0000-0001-9792-6528
[2] Faculty of Transportand
Aviation Engineering, The Silesian University of
Technology, Krasińskiego 8 Street, 40-019
Katowice, Poland. Email: marcin.j.klos@polsl.pl.
ORCID: https://orcid.org/0000-0002-4990-1593