=Paper=
{{Paper
|id=Vol-2280/paper-09
|storemode=property
|title=Comparison of Support Vector Machines and Deep Learning for Vehicle Detection
|pdfUrl=https://ceur-ws.org/Vol-2280/paper-09.pdf
|volume=Vol-2280
|authors=Özgür Kaplan,Ediz Saykol
|dblpUrl=https://dblp.org/rec/conf/rtacsit/KaplanS18
}}
==Comparison of Support Vector Machines and Deep Learning for Vehicle Detection==
https://ceur-ws.org/Vol-2280/paper-09.pdf
Comparison of Support Vector Machines and Deep Learning for Vehicle
Detection
Özgür Kaplan and Ediz Şaykol
Beykent University Department of Computer Engineering,
Ayazağa Campus, 34396, Istanbul, Turkey
ozgurkaplan@outlook.com , ediz.saykol@beykent.edu.tr
Abstract Besides, there are also simulation-based approaches
to help drivers. For example, [PYL+15] proposed an
The main goal of this paper is to compare increased reality system to increase the driver's
different Vehicle Detection algorithms and to immediate attention. The system calculates the
provide an effective comparison technique for probability of collision with the vehicles in the same
developers and researchers. During this study, lane as the driver and colorizes the lanes according to
fine tunings are suggested to improve the the risk ratio. In [RSL+14], it is aimed to help middle-
implementations of these algorithms. Our focus aged and older drivers to make a left turn in running
on Support Vector Machines (SVM) and then traffic. Drivers have been given hints to return with
Deep Learning based approaches. The SVM augmented reality for their left turns in different
based vehicle detection implementation utilizes scenarios such as heavy traffic and flowing traffic.
Histogram Oriented Gradients (HOG). The In this paper, we focus on SVM-based vehicle
deep learning approach we consider is the detection technique, where we can simplify vehicle
YOLO implementation. Our evaluation detection problem to a classification problem by
employs 400 random frames extracted from a examining individual sections of image and classify
real world driving video. As stated by the whether it is a vehicle image or not a vehicle. On the
experimental results, YOLO is more accurate other side, we consider Deep Learning approaches that
with %81.9 success than SVM which only automatically learn image features that are required for
scored %57.8. vehicle detection. There are different deep learning
techniques such as Region-based Convolutional
1. Introduction Network (R-CNN) [GDD+14], Fast R-CNN [Gir15],
Faster R-CNN [RHG+17] and You Only Look Once
According to the World Health Organization (WHO), (YOLO) [RDG+16]. We choose YOLO as it is the faster
1.25 million people die each year from traffic accidents among these techniques [DU18]. The basic aim is to
[WHO15]. Prevention of traffic accidents and provide an effective comparison technique for
punishment of drivers who violate the rules are of great developers and researchers.
importance for humanity. Many municipalities are using The organization of the paper is as follows: In Section
automated camera system to detect cars that violates 2, we present vehicle detection techniques that we
traffic rules. Proper operation of these systems is crucial employ in our study. Section 3 provides the comparative
to avoid further traffic accidents. study and Section 4 concludes the paper.
It is estimated that 1 in every 10 cars in traffic will be
composed of vehicles capable of self-driving in 2030
and 1 in every 3 in 2050 [MKG+16]. It is obvious that
2. Vehicle Detection Approaches
the safety of future traffic will be directly linked to the Our study contains two major components. First
quality of the software these vehicles will be using. For phase is detecting vehicles using SVM and deep
both software, detecting vehicles who violate the rules learning with YOLO. Second phase is comparing
and self-driving vehicles, it is critical to identify images accuracy and performance of these two methods.
correctly. As our starting point we utilized implementation of
There are several techniques in the literature suited [Fu17] that includes a classical SVM approach using
for vehicle detection task. Existing papers suggests OpenCV and Histogram of Oriented Gradients (HOG)
using acoustic sensor networks [WCH17], wavelet and feature extraction as well as a deep neural network with
interest point based feature extraction [KB13], edge- YOLO tensor flow implementation using pre-trained
based constraint filters based vehicle segmentation weight object [Cho16]. To simplify the problem we only
[Sri02] and time spatial data [MRR10]. considered cars as vehicles.
2.1 SVM-based Vehicle Detection using HOG The HOG technique counts the occurrences of the
gradient orientation of the localized regions of an image.
In this approach, we used supervised learning with First image is divided into small connected cells, and for
pre-categorized images. We used the images provided each cell gradient directions are calculated. Each cell is
by GTI vehicle database [GTI18]. There are 3425 splitted into angular bins according to the gradient
images of vehicle rears and 3900 images of road orientation. The pixel of each cell adds the weight
sequences not containing vehicles. All images are 64x64 gradient to the corresponding angular bin. Groups of
pixels. Figure 1 shows a vehicle image and a non-vehicle adjacent cells are called blocks. The grouping of cells to
image. blocks is the basis of normalization of histograms. This
normalized group of histograms generates the block
histogram and set of block histograms represents the
descriptor [Int18]. Figure 3 shows a car image and the
corresponding HOG transformation.
Figure 1: A vehicle image (left) and a non-vehicle
image (right) in [GTI18].
First step to prepare the images for SVM classifier is
to extract HOG features. The main purpose of HOG is
to identify the image as a group of local histograms.
HOG is not scale invariant. In order to use HOG we
needed the same size for our training images. All our
training set images were already the same size of 64x64
so we were able to use our images directly. Figure 2
shows our HOG implementation details.
Figure 3: HOG Transformation of a car image.
To train the model we randomly split the data into
two parts with respect to 5-fold. %80 percent is used for
training and %20 percent is used for test, as suggested
by most of the learning schemes.
We explored different parameters and color spaces
for our images to get better results. After some trial and
error we managed to get %86.7 success rate.
We wanted to test the classifier on a large scale
Figure 2: HOG Implementation. (The image source: images. We found raw driving video data on KITTI
https://software.intel.com/en-us/ipp-dev-reference- Vision Benchmark Suite [KIT18]. We modified the
histogram-of-oriented-gradients-hog-descriptor) code to make it work on 1392 x 512 pixels which is the
resolution of raw driving video data.
The 64x64 pixel block of search windows is used to estimates the B bounding boxes and C class
search the entire frame. We used 50% percent overlap probabilities. The bounding box has 5 components: x, y,
for search windows. For each window, our SVM w, h and confidence. The (x, y) are center coordinates of
classifier is used to detect if there is a vehicle or not. the box. The (w, h) are dimensions. The confidence
Figure 4 shows how search windows are used. score tells us if there is any shape in the box. If the score
is zero then there should be no object in the cell. Each
grid cell makes B predictions, so there are total of S x S
x B x 5 outputs. The network predicts one class
probability per cell that results S x S x C total
probabilities. Figure 6 shows YOLO object detection
and Figure 7 shows YOLO execution pipeline.
Figure 4: Search Windows (The image source:
https://raw.githubusercontent.com/JunshengFu/vehicle-
detection/master/examples/search_windows.png )
To get the exact locations of vehicles, heat map
generation is used with the windows that may have
vehicle. Figure 5 shows a successfully detected vehicle
using SVM.
Figure 6: YOLO Object Detection [RDG+16]
Figure 7: YOLO Pipeline [RDG+16]
Figure 5: Vehicle Detected with SVM We utilized [Fu17] code, which employs tensorflow
implementation of YOLO [Cho16]. It uses pre-trained
YOLO_small network, which has 20 classes as follows:
2.2 Deep Learning based Vehicle Detection "aeroplane", "bicycle", "bird", "boat", "bottle", "bus",
"car", "cat", "chair", "cow", "dining table", "dog",
We chose the YOLO technique for this purpose. "horse", "motorbike", "person", "potted plant", "sheep",
YOLO uses deep neural network to detect objects. Yolo "sofa", "train","tv monitor". Since vehicle is already
approaches object detection as a simple single known as “car”, we were able to use precomputed
regression problem to find bounding boxes class settings and apply it directly to our inputs. We used 30%
probabilities. YOLO is trained on full images and it threshold, cells whose car class score is 0.3 or more are
predicts multiple bounding boxes using single neural selected. Figure 8 shows a successfully detected vehicle
network. First, YOLO model accepts an image as input using YOLO.
then divides it into an SxS grid. Each cell of this grid
Figure 10 shows a sample snapshot of our
experimental study employing deep learning based
YOLO technique. YOLO downscales images to
448x448 which causes distortions. With pre road or line
detection, we can send smaller and more useful
segments to YOLO to avoid distortions and get better
results.
Figure 8: Vehicle Detected with YOLO
3. Comparing SVM and YOLO
We tested both algorithms on the same randomly
selected raw driving video data. Random data contains
400 images with total of 266 vehicles in them.
Test results are evaluated in 3 categories; Figure 10: Sample Snapshot of YOLO technique.
Positive : Vehicle detected correctly
Negative: Vehicle is not detected. Both algorithms failed to detect vehicles, which are
False Positive: Non-vehicle is detected as further ahead. We may be able to detect these further
vehicle. vehicles with SVM using higher resolution data.
However higher resolution data would require more
Figure 9 shows a sample snapshot of our computation and therefore the performance would be
experimental study employing SVM-based technique. worse. As for YOLO, which downscales pictures to
SVM generated a lot of false positives. In some frames, 448x448 pixels, it is useless to use higher resolution
SVM identified road signs, trees and pedestrians as data.
vehicles. These false negatives may be avoided if a pre
road line detection is performed. Thus, we can limit
SVM search window within actual road.
Figure 11: Test results for SVM
Figure 9: Sample Snapshot of SVM-based technique.
4. Conclusion and Future Work
In this paper, we looked into two different vehicle
detection algorithm implementations and test them
against real world traffic data.
Our results showed that deep neural network with
YOLO performed more accurate results. YOLO also
performed faster which makes it more suited for real
time tasks.
With our approach we are able to find strong and
weak sides of both models. Thus, we managed to
implement and suggest fine tunings. Our approach has
potential to be used to compare different algorithms and
to be used to find fine tunings.
Figure 12: Test results for YOLO
References
[WHO15] World Health Organization. (2015). Global
status report on road safety 2015. World
Health Organization.
http://www.who.int/iris/handle/10665/189242
[MKG+16] D.Mohr, H.Kaas, P.Gao, D.Wee, and
T.Möller, “Automotive revolution—
Perspective towards 2030: How the
convergence of disruptive technology-driven
trends could transform the auto industry,”
McKinsey & Company, Washington, DC,
USA, Tech. Rep., Jan. 2016
[WCH17]Wang, R., Cao, W., & He, Z. (2017). Vehicle
Figure 13: Overall Test Results. recognition in acoustic sensor networks using
multiple kernel sparse representation over
Our test results shows that YOLO’s success rate is learned dictionaries. International Journal of
81.9%, 218 out of 266, and SVM’s is %57.8, 154 out of Distributed Sensor Networks.
266. It also has significantly lower false positives with 5 https://doi.org/10.1177/1550147717701435
against SVM’s 84. When we apply false negatives [KB13] KumarMishra, Pradeep & Banerjee, Biplab.
YOLO’s success rate drops to 80.4% and SVM’s drops (2013). Multiple Kernel based KNN
to 44%. Classifiers for Vehicle Classification.
One reason of high false positives of SVM is that
International Journal of Computer
unlike YOLO, who can classify 20 different types, SVM
Applications. 71. 1-7. 10.5120/12359-8673.
can only identify vehicles. YOLO lowers the class
probability if there is another possible object in the [Sri02] Srinivasa, Narayan. (2002). Vision-based
frame, which lowers the chance of false positives. vehicle detection and tracking method for
We processed the entire video using Asus Nvdia forward collision warning in automobiles.
Geforce 1060 gpu. YOLO performed up to 42 fps while 10.1109/IVS.2002.1188021.
SVM reach just 4 fps. This is because YOLO is [MRR10] N. C. Mithun, N. U. Rashid and S. M. M.
lightweight and use single neural network on given Rahman, "Detection and Classification of
frame. On the other hand, SVM needs recursive Vehicles From Video Using Multiple Time-
calculations with multiple sliding window calculations. Spatial Images," in IEEE Transactions on
Intelligent Transportation Systems, vol. 13,
no. 3, pp. 1215-1225, Sept. 2012. doi: [Cho16] Jinyoung Choi. (2016). Tensorflow
10.1109/TITS.2012.2186128 Implementation of ‘YOLO: Real-Time Object
[PYL+15] B. J. Park, C. Yoon, J. W. Lee and K. H. Kim, Detection’. Retrieved from
"Augmented reality based on driving situation https://github.com/gliese581gg/YOLO_tensorf
awareness in vehicle," 2015 17th International low
Conference on Advanced Communication [GTI18] Vehicle Image Database (2018, May 12)
Technology (ICACT), Seoul, 2015, pp. 593- http://www.gti.ssr.upm.es/data/Vehicle_databa
595. doi: 10.1109/ICACT.2015.7224865 se.html
[RSL+14] Rusch, Michelle & Schall, Mark & Lee, John [Int18] Histogram of Oriented Gradients (HOG)
& Dawson, Jeffrey & Rizzo, Matthew. (2014). Descriptor (2018, April 2)
Augmented reality cues to assist older drivers https://software.intel.com/en-us/ipp-dev-
with gap estimation for left-turns. Accident reference-histogram-of-oriented-gradients-
Analysis & Prevention. 71. 210–221. hog-descriptor
10.1016/j.aap.2014.05.020. [KIT18] KITTI Vision Bench. Suite (2018, May 13)
[GDD+14] R. Girshick, J. Donahue, T. Darrell and J. http://www.cvlibs.net/datasets/kitti/raw_data.p
Malik, "Rich Feature Hierarchies for Accurate hp
Object Detection and Semantic Segmentation,"
2014 IEEE Conference on Computer Vision
and Pattern Recognition, Columbus, OH, 2014,
pp. 580-587.doi: 10.1109/CVPR.2014.81
[Gir15] R. Girshick, "Fast R-CNN," 2015 IEEE
International Conference on Computer Vision
(ICCV), Santiago, 2015, pp. 1440-1448. doi:
10.1109/ICCV.2015.169
[RHG+17] S. Ren, K. He, R. Girshick and J. Sun,
"Faster R-CNN: Towards Real-Time Object
Detection with Region Proposal Networks," in
IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 39, no. 6, pp. 1137-
1149, 1 June 2017. doi:
10.1109/TPAMI.2016.2577031
[RDG+16] J. Redmon, S. Divvala, R. Girshick and A.
Farhadi, "You Only Look Once: Unified,
Real-Time Object Detection," 2016 IEEE
Conference on Computer Vision and Pattern
Recognition (CVPR), Las Vegas, NV, 2016,
pp. 779-788
[Du18] Juan Du (2018). Understanding of Object
Detection Based on CNN Family and YOLO.
Journal of Physics: Conference Series, 1004,
012029.
[Fu17] Juhnsheng FU. (2017, March 17). Vehicle
Detection for Autonomous Driving. Retrieved
from https://github.com/JunshengFu/vehicle-
detection