This is a postprint version of the published document at:

Espinosa, J.E., Velastin, S.A. y Branch, J.W. (2017).

Vehicle Detection Using Alex Net and Faster R-CNN Deep
Learning Models: A Comparative Study. In Advances in
Visual Informatics. Lecture Notes in Computer Science,
10645, pp. 3-15.

DOI: https://doi.org/10.1007/978-3-319-70010-6_1

© Springer Nature Switzerland AG 2017.

 Vehicle Detection Using Alex Net and Faster
R-CNN Deep Learning Models:
A Comparative Study

Jorge E. Espinosa1, Sergio A. Velastin2,3(&), and John W. Branch4

1
Facultad de Ingenierías,
Politécnico Colombiano Jaime Isaza Cadavid – Medellín, Medellín, Colombia
[email protected]
2
University Carlos III - Madrid, Madrid, Spain
[email protected]
3
Queen Mary University of London, London, UK
4
Facultad de Minas, Universidad Nacional de Colombia – Sede Medellín,
Medellín, Colombia
[email protected]

Abstract. This paper presents a comparative study of two deep learning models
used here for vehicle detection. Alex Net and Faster R-CNN are compared with
the analysis of an urban video sequence. Several tests were carried to evaluate
the quality of detections, failure rates and times employed to complete the
detection task. The results allow to obtain important conclusions regarding the
architectures and strategies used for implementing such network for the task of
video detection, encouraging future research in this topic.

Keywords: Convolutional Neural Network
Feature extraction
Vehicle

classification

1 Introduction

Currently traffic management is supported by urban traffic analysis. Traditionally,

vehicle counting and road density evaluations are done with inductive loop sensors and
increasingly with video information. Nevertheless, video detection faces different
challenges due to changes in illumination conditions and high vehicle densities with
frequent occlusions. Most video detection systems are based on appearance features or
motion features. Appearance features [1–4] such as shape, color, edge maps and tex-
ture, are used to detect vehicles even in stationary positions. Other works are based on
HOG (Histogram of Oriented Gradients) and some variations of it [5, 6]. Motion
features are obtained based on the dynamics of the traffic movement. Such methods are
generally based on background subtraction [7], use of frame difference [8], Kalman filter
[9], optical flow [10, 11], etc. For a detailed survey of traditional vehicle detection
methods please see [12].
On the other hand, deep learning theory (DL) applied to image processing is the
current dominant computer vision theory especially in tasks such as image recognition.

1
Since 2010 the annual image recognition challenge known as the ImageNet
Large-Scale Visual Recognition Competition (ILSVRC) [13] is being dominated by
this approach.
For vehicle detection, several works using Deep Learning in vehicle detection are
reported in the literature. Earlier approaches relied on 2D Deep Belief Networks (2D-
DBN) [14], learning features by means of this architecture and using a pre-training
sparse filtering process [15] or Hybrid architectures (HDNN) which overcome the issue
of the single scale extraction features of traditional DNNs [16]. Color as a discriminative
feature is used in [17] and [18]. There are also pre-training schemes [19] that obtain
competitive results even with low resolution images and implementable in real time as in
[20]. More recently, detection and classification of multiples classes is performed using
integrated models as Fast R-CNN and Faster R-CNN [21–26]. Reports exist of methods
able to recognize vehicle make and models (MMR) [27, 28], re-identification archi-
tectures for security urban surveillance [29–31], strategies using DBN [14, 32–34] that
work with relatively few labelled data and models that are able to classify even the pose
or orientation of the vehicle [23, 35, 36]. Generally most of the detection and classifi-
cations models are implemented using different CNN architectures such as CaffeNet
[37, 38], GoogLeNet [39], and VGGNet [26] used in [27]. AlexNet [40] is used by Su et
al.[18] in conjunction with GoogleNet [39] and NIN (Network in Network) [41].
Nevertheless, as far as we know, there are no comparative studies of DL strategies
used for vehicle classification, nor on the use of CNNs already trained for feature
extraction to perform vehicle discrimination in video sequences.
This work compares the results of a CNN used for feature extraction and a CNN
integrated model network, both used for the task of classifying vehicles in video
sequences. The paper is organized as follows: Sect. 2 gives a brief explanation of the
architecture of the convolutional neural networks, explaining the advantage of the use of
an already-trained network for feature extraction and the benefits of the integrated CNN
model. Section 3 shows the classification approaches, describing the character-istics of
the models built for the video detection task. Section 4 shows the results of the two
models, comparing and explaining the results. Section 5 presents the conclusions and
proposes some future work.

2 CNN Architectures Used

In this section, we describe the principal characteristics of the CNN AlexNet and the
Faster R-CNN networks used in this comparative study.

2.1 AlexNet
AlexNet is considered the pioneer work of CNN networks, even after the work of Yann
LeCun [42]. The AlexNet model was introduced in the paper “ImageNet Classification
with Deep Convolutional Networks”, were the authors created a “large, deep convo-
lutional neural network”, used to win the 2012 ILSVRC (ImageNet Large-Scale Visual
Recognition Challenge) [43]. The network was trained on ImageNet data, with over 15
million annotated images from a total of over 22,000 categories.

2
 The architecture of “AlexNet” has 23 layers, integrating 5 convolution layers, 5
ReLu layers (Rectified units), 2 layers for normalization, 3 pooling layers, 3 fully
connected layers, one probabilistic layer with softmax units and finally a classification
layer ending in 1000 neurons for 1000 categories (Fig. 1).

Fig. 1. Alex Net architecture. [40]

The main characteristics of the network includes the use of ReLU for the nonlin-
earity functions that decrease the training time as ReLUs are faster than the conven-
tional tanh function used in MLP. For training proposes, the authors used techniques as
data augmentation consisting in horizontal reflections, image translations or even patch
extractions. Dropout layers were also included to overcome the problems of vanishing
gradient and overfitting. The model was trained using batch stochastic gradient descent,
using specific values for momentum and weight decay. It took nearly six days for
training using two GTX 580 GPUs.

2.2 Faster R-CNN

The problem of object detection involves the detection of different classes of objects in
each image. Traditionally, the strategy used was to deploy a two-class classifier (object
vs non-object) in conjunction with a sliding window search. The amount of all win-
dows scales and ratios returned could be huge, and then it was necessary to implement
methods like non-maximal suppression for reducing the redundant candidates. This was
the origin of object proposal algorithms, with strategies to limit search using calibration
information [44], Branch & Bound [45] or grouping adjacent pixels merging them to
find a blob region as in Selective Search [46]. These methods also included pre defining
windows based on objects candidates as in Spatial Pyramid Pooling [47], or Edge boxes
[48]. A valuable comparison of such methods was done by Hosang et al. [49]. The pre-
filtering strategy has been used with positive results combining it with CNN networks
for classification as in R-CNN [50] but employing too much time in the training process.
To improve the training process time, Fast R-CNN [21] has been proposed that swaps
the extracting strategy of detecting regions and running CNN. A high-resolution image
is fed to the CNN network, the network produces a high resolution convolutional feature
map. The region proposal produces regions over the feature map (conv5). The
convolutional features of these regions are

3
then fed into fully connected layers, with a linear classifier and a bounding box linear
regression module to define regions. This model continues slowly at test time. Faster
R-CNN addressed this issue by combining features of a fully convolutional network to
perform both region proposals and object detection. Since region proposals depended
on features of the image that were already calculated with the forward pass of the CNN
(first step of classification), the model reuses the same CNN results for region proposals
instead of running a separate selective search algorithm. The region proposal network
(RPN) shares convolutional layers with the object detection network, then only one
CNN needs to be trained and region proposals is calculated almost for free. Then,
additional convolutional layers are used to regress region bounds with scores for object
proposal at each location. The RPN works by moving a sliding window over the CNN
feature map and at each window, generating k potential bounding boxes and scores
associated for how good each of those boxes is expected to be. This k represents the
common aspect ratios that candidates to objects could fit, caller anchor boxes. For each
anchor box, the RPN output a bounding box and score per position in the image. This
model improves significantly the speed and the object detection results.
Besides achieving the highest accuracy on both PASCAL VOC 2007 and 2012,
Faster R-CNN was the basis of more than 125 proposed entries in ImageNet detection
and localization at ILSVRC 2016 [51] and in the COCO challenge 2015 was the
foundation of the winners in several categories [25]. Figure 2 shows the network
structure of the Faster R-CNN framework. Both the region proposal network and the
object classifier share fully convolutional layers. These layers are trained jointly. The
region proposal network behaves as an attention director, determining the optimal
bounding boxes across a wide range of scales and using nine candidate aspect ratios to
be evaluated for object classi-fication. In other words, the RPN tells the unified network
where to look.

Fig. 2. Faster R-CNN network structure. Modified from [24]

3 Detection and Classification Approaches

An initial ROI (region of interest) step is implemented, allowing the user to select the
precise area of analysis. This step optimizes performance and speeds up the processes
of detection and classification, by reducing the area of analysis in the video sequences.
Two models, AlexNet and Faster R-CNN were used for vehicle detection and classi-
fication in this research.

4
3.1 AlexNet Model
In the AlexNet model, object detection is performed based on background subtraction,
using a Gaussian Mixture Model according to Zivkovic [52, 53]. Once the objects are
detected, these are classified as vehicles on three categories: motorcycles, cars or buses.
Other possible objects detected are classified as part of the urban environment (“urb-
Tree”), any possible remaining detection is assigned to the “unknown” class. The
classifier is constructed using a multiclass linear SVM trained with features obtained
from a pre-trained CNN. Those features have been extracted for a set of 80 images per
category, including the “urbTree” category created from the urban traffic environment.
This approach to image category classification is based on the work published by Matlab
in “Image category classification using deep learning” [54] and extended in [55] for
Motorcycles classification. Here, we extend the categories to include the bus set. The
classifier trained using CNN features provides close to 100% accuracy, which is higher
than the accuracy achieved using methods such as Bag of Features and SURF. As in [55]
the set of images categories has been created corresponding to images related to
motorcycles, cars, buses and urban environment related objects (“urbTree”). Those
images were obtained from different angles and perspectives in urban traffic in
Medellin City (Colombia). Examples of each category are shown in Fig. 3.

Fig. 3. Examples of cars, motorbikes and buses.

Following the strategy described in [55] by using the selected images, the
pre-trained CNN “AlexNet” network is used for feature extraction, this technique is
detailed described by Razavian et al. in [56]. For this work, the AlexNet network is
only used to classify four categories. This pre-trained network was used to learn
motorcycles, cars and buses features obtained from the extended dataset, with 80
images per category and 80 examples of the class “urbTree” created from the urban
environment. In the end, the total number of examples is only 320.
Features are extracted from the training set, propagating images through the net-
work up to a specific fully connected layer (fc7), extracting activations responses to
create a training set of features, which is used later for classification.
For classification, as in [55], a multiclass SVM classifier is trained using the image
features obtained from the CNN. Since the length of the feature vector is 4096, a fast
stochastic gradient descent solver is used as training algorithm. In this case the clas-sifier
is trained with only 96 examples (24 per category). The validation set, which

5
corresponds to the remaining 224 examples (56 by category) is then classified. The
classifier accuracy is evaluated now with the features obtained on this set. Figure 4
shows the results in a confusion matrix. The classifier mismatches three bus images
classifying those as cars; one car image is classified as bus and another as a motorcycle.
The mean accuracy obtained is 0.978.

Fig. 4. Confusion matrix of the experiments (Class 1: Buses 2: Cars 3: Motorcycles 4: urbTree)

3.2 Faster R-CNN Model

The Faster R-CNN [25] model used in this research, is the one available for download
at https://github.com/ShaoqingRen/faster_rcnn. The model is an object detection
framework based on deep convolutional neural networks, which includes two net-
works: A Region Proposal Network (RPN) and an Object Detection Network. Both
networks are trained for sharing convolutional layers to obtain real time results.
The model can be run based on two referenced CNN networks: ZF Net [57] or
VGG16 [26]. In this research, we chose VGG16, which corresponds to the best per-
formance results given the network layer configuration that the literature reports on the
use of this architecture.
The VGG16 based model is pretrained with the ImageNet dataset. After down-
loading, both networks (RPN and ODN) are retrained in the PASCAL VOC 2007
dataset (Fig. 5) [58]. Based on the dataset used for retraining, the ODN is able to
classify detections on 20 categories as follows:
• Person: person
• Animal: bird, cat, cow, dog, horse, sheep
• Vehicle: aeroplane, bicycle, boat, bus, car, motorbike, train
• Indoor: bottle, chair, dining table, potted plant, sofa, tv/monitor

6
 Fig. 5. Examples of each category in PASCAL VOC 2007 dataset [58].

For this experiments all classes different to bus, car or motorbike, are renamed as
“unknown” to obtain comparative metrics for the results evaluation. As a preliminary
step for video detection and classification, the user defines a ROI within which the
detection and classification takes place.

4 Experiments and Results

We selected a video sequence of a real urban environment took in Medellín Colombia,

over a secondary street of two lanes. The sequence consists of 1812 RGB frames
(640  480), during daylight and good weather conditions. The sequence includes 36
different cars to detect (including sedan, VAN or taxis), 7 motorbikes and 1 Bus, which
is the class assigned to a detected truck.
An XML-coded ground truth was obtained by means of ViPER [59] tool for
annotation, and is used to compare the results of the two models. We use the perfor-
mance metrics reported in [60]. These metrics have been extended to take into account
the multiclass nature of the experiment.
The evaluations for Faster R-CNN are performed based on the NMS parameter
threshold, used to reduce the redundancy on proposed regions. This threshold corre-
sponds to the IoU overlap of the proposed regions. Results described in Table 1 and
Fig. 6, show that decreasing the IoU threshold criteria increases the correct detection
rate in total and for each class analyzed, but at the same time increases the False Alarm
Rate. Best results correspond to a NMS threshold of 0.6 with a F1-Score of 0.76.

Table 1. Rates of Faster R-CNN results. NMS (non maximal suppression) – CDR: correct
detection rate, Bikes: CDR for motorcycles, Cars: CDR for cars, Bus: CDR for buses or trucks.
DFR: Detection Failure Rate. FAR: False Alarm Rate. PR: Precision. RC: Recall. F1:F1-score.
NMS threshold CDR CDR bikes CDR CDR buses DFR FAR PR RC F1
cars
0.30 0.75 0.26 0.83 0.33 0.25 0.38 0.62 0.75 0.68
0.40 0.73 0.21 0.81 0.32 0.27 0.31 0.69 0.73 0.71
0.50 0.72 0.18 0.80 0.33 0.28 0.22 0.78 0.72 0.75
0.60 0.70 0.13 0.78 0.29 0.30 0.16 0.84 0.70 0.76
0.70 0.65 0.09 0.74 0.23 0.35 0.13 0.87 0.65 0.75
0.80 0.61 0.06 0.70 0.07 0.39 0.10 0.90 0.61 0.73

7
 Fig. 6. Faster R-CNN rates results.

Meanwhile, working with the AlexNet classifier in conjunction with the GMM
background subtraction, results are obtained in terms of the parameters of the back-
ground subtraction algorithm. First, the history parameter is evaluated against a fixed
Mahalanobis distance of 128. History corresponds to the number of frames (LoH) that
constitutes the training set for the background model. The best results obtained are for a
history of 500 frames (F1 = 0.57). Fixing this number, we then proceed to change the
Mahalanobis distance parameter (Tg). This parameter is a threshold for the squared
Mahalanobis distance that helps decide when a sample is close to the existing com-
ponents. A smaller Tg value generates more components. A higher Tg value may result
in a small number of components but they can grow too large. The best result is
obtained with LoH of 500, and Tg of 20, achieving a CDR of 0.66 with a FAR of 0.32
(Fig. 7 and Tables 2, 3).

Fig. 7. Alex Net rates results

8
Table 2. Rates of Faster R-CNN results. NMS (non maximal suppression) – CDR: correct
detection rate, Bikes: CDR for motorcycles, Cars: CDR for cars, Bus: CDR for buses or trucks.
DFR: Detection Failure Rate. FAR: False Alarm Rate. MR: Merge Rate. PR: Precision. RC:
Recall. F1:F1-score.
LoH CDR CDR bikes CDR CDR buses DFR FAR MR PR RC F1
cars
100 0.32 0.45 0.29 0.24 0.66 0.43 0.02 0.57 0.34 0.41
200 0.46 0.60 0.41 0.47 0.51 0.36 0.02 0.64 0.49 0.54
300 0.50 0.57 0.45 0.68 0.47 0.33 0.01 0.67 0.53 0.57
400 0.51 0.53 0.46 0.68 0.46 0.36 0.01 0.64 0.54 0.57
500 0.52 0.53 0.47 0.66 0.44 0.36 0.01 0.64 0.56 0.58
600 0.51 0.51 0.47 0.67 0.45 0.38 0.01 0.62 0.55 0.56

Table 3. Rates of Faster R-CNN results. NMS (non maximal suppression) – CDR: correct
detection rate, Bikes: CDR for motorcycles, Cars: CDR for cars, Bus: CDR for buses or trucks.
DFR: Detection Failure Rate. FAR: False Alarm Rate. MR: Merge Rate. PR: Precision. RC:
Recall. F1:F1-score.
Tg CDR CDR bikes CDR CDR buses DFR FAR MR PR RC F1
cars
16 0.64 0.50 0.66 0.22 0.36 0.42 0.08 0.58 0.64 0.61
20 0.66 0.50 0.68 0.24 0.33 0.38 0.08 0.62 0.67 0.64
32 0.68 0.53 0.69 0.29 0.31 0.33 0.07 0.67 0.69 0.67
40 0.67 0.53 0.67 0.32 0.32 0.31 0.06 0.69 0.68 0.68
48 0.66 0.54 0.65 0.37 0.33 0.31 0.06 0.69 0.67 0.67
64 0.64 0.55 0.62 0.45 0.35 0.30 0.05 0.70 0.65 0.66

The results obtained show that Faster R-CNN outperforms Alex Net+GMM model,
not only in the correct detection rate obtained while producing less false detections, but
also in time spending in the analysis. Both model were analyzed on a Windows 10
Machine with a core i7 7th generation, 4.7 GHz, and 32 GB of RAM, using an NVidia
Titan X (Pascal) 1531 MHz GPU, achieving close to real time in Faster R-CNN model
(40 ms per frame) while AlexNet+GMM took almost 100 ms per frame.

5 Conclusions and Future Work

This paper has compared the performance of two deep learning models for vehicle
detection and classification in urban video sequences. Although the AlexNet model is
used for feature extraction in an ad-hoc set of examples oriented to urban scenarios, the
pre-trained Faster R-CNN model achieves better results in correct detections according
to F1-score measure. It is important to remark that the Faster R-CNN model does not
use any dynamic attributes for vehicle detection whereas GMM background subtraction
used in AlexNet model. In fact, as the merge rates (MR) result shows, GMM

9
background subtraction still has issues with stationary vehicles and occluded scenarios.
In Faster R-CNN, the RPN component results could be improved providing some
urban context information as restriction size of the regions.
For future work, we intend to improve the results of the RPN component of the
Faster R-CNN model enriching it with traffic context information, and improve the
classification component with feature extraction using a Deep Architecture as AlexNet,
ZF or VGG, with a wider set of urban road user classes (e.g. trucks, vans, cyclists,
pedestrians).

Acknowledgments. S.A. Velastin is grateful to funding received from the Universidad Car-
los III de Madrid, the European Union’s Seventh Framework Programme for research, techno-
logical development and demonstration under grant agreement no. 600371, el Ministerio de
Economía y Competitividad (COFUND2013-51509) and Banco Santander. The authors wish to
thank Dr. Fei Yin for the code for metrics employed for evaluations. Finally, we gratefully
acknowledge the support of NVIDIA Corporation with the donation of the GPUs used for this
research. The data and code used for this work is available upon request from the authors.

References
1. Tsai, L.W., Hsieh, J.W., Fan, K.C.: Vehicle detection using normalized color and edge
map. IEEE Trans. Image Process. 16(3), 850–864 (2007)
2. Ma, X., Grimson, W.E.L.: Edge-based rich representation for vehicle classification. In: 10th
IEEE International Conference on Computer Vision (ICCV 2005), vol. 1–2, pp. 1185–1192
(2005)
3. Buch, N., Orwell, J., Velastin, S.A.: 3D extended histogram of oriented gradients (3DHOG)
for classification of road users in urban scenes (2009)
4. Feris, R.S., et al.: Large-scale vehicle detection, indexing, and search in urban surveillance
videos. IEEE Trans. Multimed. 14(1), 28–42 (2012)
5. Chen, Z., Ellis, T.: Multi-shape descriptor vehicle classification for urban traffic. In: 2011
International Conference on Digital Image Computing Techniques and Applications
(DICTA), pp. 456–461 (2011)
6. Chen, Z., Ellis, T., Velastin, S.A.: Vehicle detection, tracking and classification in urban
traffic. In: 2012 15th International IEEE Conference on Intelligent Transportation Systems,
pp. 951–956 (2012)
7. Gupte, S., Masoud, O., Martin, R.F., Papanikolopoulos, N.P.: Detection and classification of
vehicles. IEEE Trans. Intell. Transp. Syst. 3(1), 37–47 (2002)
8. Cucchiara, R., Piccardi, M., Mello, P.: Image analysis and rule-based reasoning for a traffic
monitoring system. IEEE Trans. Intell. Transp. Syst. 1(2), 119–130 (2000)
9. Messelodi, S., Modena, C.M., Zanin, M.: A computer vision system for the detection and
classification of vehicles at urban road intersections. Pattern Anal. Appl. 8(1–2), 17–31
(2005)
10. Huang, C.-L., Liao, W.-C.: A vision-based vehicle identification system. In: Proceedings of
17th International Conference on Pattern Recognition, ICPR 2004, vol. 4, pp. 364–367
(2004)
11. Ottlik, A., Nagel, H.-H.: Initialization of model-based vehicle tracking in video sequences of
inner-city intersections. Int. J. Comput. Vis. 80(2), 211–225 (2008)

10
12. Tian, B., et al.: Hierarchical and networked vehicle surveillance in its: a survey. IEEE Trans.
Intell. Transp. Syst. 16(2), 557–580 (2015)
13. ImageNet Large Scale Visual Recognition Competition (ILSVRC). http://www.image-net.
org/challenges/LSVRC/. Accessed 24 Oct 2016
14. Wang, H., Cai, Y., Chen, L.: A vehicle detection algorithm based on deep belief network.
Sci. World J. 2014, e647380 (2014)
15. Dong, Z., Pei, M., He, Y., Liu, T., Dong, Y., Jia, Y.: Vehicle type classification using
unsupervised convolutional neural network. In: 2014 22nd International Conference on
Pattern Recognition (ICPR), pp. 172–177 (2014)
16. Chen, X., Xiang, S., Liu, C.L., Pan, C.H.: Vehicle detection in satellite images by hybrid
deep convolutional neural networks. IEEE Geosci. Remote Sens. Lett. 11(10), 1797–1801
(2014)
17. Hu, C., Bai, X., Qi, L., Chen, P., Xue, G., Mei, L.: Vehicle color recognition with spatial
pyramid deep learning. IEEE Trans. Intell. Transp. Syst. 16(5), 2925–2934 (2015)
18. Su, B., Shao, J., Zhou, J., Zhang, X., Mei, L.: Vehicle color recognition in the surveillance
with deep convolutional neural networks (2015)
19. Zhang, F., Xu, X., Qiao, Y.: Deep classification of vehicle makers and models: the
effectiveness of pre-training and data enhancement. In: 2015 IEEE International Conference
on Robotics and Biomimetics (ROBIO), pp. 231–236 (2015)
20. Bautista, C.M., Dy, C.A., Mañalac, M.I., Orbe, R.A., Cordel, M.: Convolutional neural
network for vehicle detection in low resolution traffic videos. In: 2016 IEEE Region 10
Symposium (TENSYMP), pp. 277–281 (2016)
21. Girshick, R.: Fast R-CNN. In: Proceedings of IEEE International Conference on Computer
Vision, pp. 1440–1448 (2015)
22. Wang, S., Liu, F., Gan, Z., Cui, Z.: Vehicle type classification via adaptive feature clustering
for traffic surveillance video. In: 2016 8th International Conference on Wireless Commu-
nications Signal Processing (WCSP), pp. 1–5 (2016)
23. Chabot, F., Chaouch, M., Rabarisoa, J., Teulière, C., Chateau, T.: Deep MANTA: a coarse-
to-fine many-task network for joint 2D and 3D vehicle analysis from monocular image. arXiv
preprint arXiv:170307570 (2017)
24. Fan, Q., Brown, L., Smith, J.: A closer look at faster R-CNN for vehicle detection. In: 2016
IEEE Intelligent Vehicles Symposium (IV), pp. 124–129 (2016)
25. Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with
region proposal networks. In: Advances in Neural Information Processing Systems, pp. 91–
99 (2015)
26. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:14091556 (2014)
27. Liu, D., Wang, Y.: Monza: image classification of vehicle make and model using
convolutional neural networks and transfer learning. http://cs231n.stanford.edu/reports/2015/
pdfs/lediurfinal.pdf. Accessed 16 Oct 2017
28. Gao, Y., Lee, H.J.: Local tiled deep networks for recognition of vehicle make and model.
Sensors 16(2), 226 (2016)
29. Liu, X., Liu, W., Mei, T., Ma, H.: A deep learning-based approach to progressive vehicle re-
identification for urban surveillance. In: European Conference on Computer Vision, pp. 869–
884 (2016)
30. Bromley, J., et al.: Signature verification using a “siamese” time delay neural network.
IJPRAI 7(4), 669–688 (1993)
31. Su, B., Shao, J., Zhou, J., Zhang, X., Mei, L., Hu, C.: The precise vehicle retrieval in traffic
surveillance with deep convolutional neural networks. Int. J. Inf. Electron. Eng. 6(3), 192
(2016)

11
32. Cai, Y., Sun, X., Wang, H., Chen, L., Jiang, H.: Night-time vehicle detection algorithm
based on visual saliency and deep learning. J. Sens. 2016 (2016)
33. Wu, Y.Y., Tsai, C.M.: Pedestrian, bike, motorcycle, and vehicle classification via deep
learning: deep belief network and small training set. In: 2016 International Conference on
Applied System Innovation (ICASI), pp. 1–4 (2016)
34. Huang, B.-J., Hsieh, J.-W., Tsai, C.-M.: Vehicle detection in Hsuehshan Tunnel using
background subtraction and deep belief network. In: Asian Conference on Intelligent
Information and Database Systems, pp. 217–226 (2017)
35. Zhou, Y., Liu, L., Shao, L., Mellor, M.: DAVE: a unified framework for fast vehicle detection
and annotation. In: Leibe, B., Matas, J., Sebe, N., Welling, M. (eds.) ECCV 2016. LNCS, vol.
9906, pp. 278–293. Springer, Cham (2016). doi:10.1007/978-3-319-46475-6_18
36. You, R., Kwon, J.-W.: VoNet: vehicle orientation classification using convolutional neural
network. In: Proceedings of 2nd International Conference on Communication and
Information Processing, pp. 195–199 (2016)
37. Caffe — Deep Learning Framework. http://caffe.berkeleyvision.org/. Accessed 05 Sept 2016
38. Luo, X., Shen, R., Hu, J., Deng, J., Hu, L., Guan, Q.: A deep convolution neural network
model for vehicle recognition and face recognition. Procedia Comput. Sci. 107, 715–720
(2017)
39. Szegedy, C., et al.: Going deeper with convolutions. In: Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition, pp. 1–9 (2015)
40. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional
neural networks. In: Advances in Neural Information Processing Systems, pp. 1097–1105
(2012)
41. Lin, M., Chen, Q., Yan, S.: Network in network. arXiv preprint arXiv:13124400 (2013)
42. Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document
recognition. Proc. IEEE 86(11), 2278–2324 (1998)
43. ImageNet Large Scale Visual Recognition Competition 2012 (ILSVRC 2012). http://www.
image-net.org/challenges/LSVRC/2012/. Accessed 30 Aug 2017
44. Brown, L.M., Fan, Q., Zhai, Y.: Self-calibration from vehicle information. In: 2015 12th
IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS),
pp. 1–6 (2015)
45. Lampert, C.H., Blaschko, M.B., Hofmann, T.: Efficient subwindow search: a branch and
bound framework for object localization. IEEE Trans. Pattern Anal. Mach. Intell. 31(12),
2129–2142 (2009)
46. Uijlings, J.R., Van De Sande, K.E., Gevers, T., Smeulders, A.W.: Selective search for object
recognition. Int. J. Comput. Vis. 104(2), 154–171 (2013)
47. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks
for visual recognition. IEEE Trans. Pattern Anal. Mach. Intell. 37(9), 1904–1916 (2015)
48. Zitnick, C.L., Dollár, P.: Edge boxes: locating object proposals from edges. In: Fleet, D.,
Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8693, pp. 391–405.
Springer, Cham (2014). doi:10.1007/978-3-319-10602-1_26
49. Hosang, J., Benenson, R., Dollár, P., Schiele, B.: What makes for effective detection
proposals? IEEE Trans. Pattern Anal. Mach. Intell. 38(4), 814–830 (2016)
50. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object
detection and semantic segmentation. In: 2014 IEEE Conference on Computer Vision and
Pattern Recognition, pp. 580–587 (2014)
51. ILSVRC2016. http://image-net.org/challenges/LSVRC/2016/results. Accessed 30 Aug 2017
52. Zivkovic, Z.: Improved adaptive Gaussian mixture model for background subtraction. In:
Proceedings of 17th International Conference on Pattern Recognition, ICPR 2004, vol. 2,
pp. 28–31 (2004)

12
53. Zivkovic, Z., Van Der Heijden, F.: Efficient adaptive density estimation per image pixel for
the task of background subtraction. Pattern Recognit. Lett. 27(7), 773–780 (2006)
54. Image Category Classification Using Deep Learning - MATLAB & Simulink Example.
https://www.mathworks.com/help/vision/examples/image-category-classification-using-
deep-learni ng.html. Accessed 28 Feb 2017
55. 8th International Conference on Pattern Recognition Systems |Universidad Carlos III de
Madrid — Madrid, Spain. http://velastin.dynu.com/icprs17/programme.php. Accessed 30
Aug 2017
56. Razavian, A.S., Azizpour, H., Sullivan, J., Carlsson, S.: CNN features off-the-shelf: an
astounding baseline for recognition. In: 2014 IEEE Conference on Computer Vision and
Pattern Recognition Workshops, pp. 512–519 (2014)
57. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Fleet,
D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8689, pp. 818–833.
Springer, Cham (2014). doi:10.1007/978-3-319-10590-1_53
58. The PASCAL Visual Object Classes Challenge 2007 (VOC 2007). http://host.robots.ox.ac.
uk/pascal/VOC/voc2007/index.html. Accessed 31 Aug 2017
59. ViPER: The Video Performance Evaluation Resource. http://viper-toolkit.sourceforge.net/.
Accessed 31 Aug 2017
60. Yin, F., Makris, D., Velastin, S.A.: Performance evaluation of object tracking algorithms. In:
IEEE International Workshop on Performance Evaluation of Tracking and Surveillance, Rio
De Janeiro, Brazil, p. 25 (2007)

13