International Journal of Science and Research (IJSR)

ISSN: 2319-7064
SJIF (2022): 7.942

Object Detection Using Python

Khusbhu Bharti
Master of Technology in Computer Science & Engineering, Lingaya‟s Vidyapeeth, Faridabad, India

Abstract: Computer Vision is the branch of the science of computers and software systems which can recognize as well as understand
images and scenes. Computer Vision is consists of various aspects such as image recognition, object detection, image generation, image
super-resolution and many more. Object detection is widely used for face detection, vehicle detection, pedestrian counting, web images,
security systems and self-driving cars. In this project, we are using highly accurate object detection-algorithms and methods such as R-
CNN, Fast-RCNN, Faster-RCNN, RetinaNet and fast yet highly accurate ones like SSD and YOLO. Using these methods and
algorithms, based on deep learning which is also based on machine learning require lots of mathematical and deep learning frameworks
understanding by using dependencies such as Tensor Flow, OpenCV, image AI etc, we can detect each and every object in image by the
area object in an highlighted rectangular boxes and identify each and every object and assign its tag to the object. This also includes the
accuracy of each method for identifying objects.

Keywords: Object detection, Computer Vision, Web images, R-CNN, Tensor Flow, OpenCV, Convolutional network, Bayes decision
rule, Gaussian distribution function

1. Introduction classifies one or more objects in an image. When a user or

practitioner refers to the term “Object Recognition”, they
A few years ago, the creation of the software and hardware often mean “Object detection”. It may be challenging for
image processing systems was mainly limited to the beginners to distinguish between different related computer
development of the user interface, which most of the vision tasks.
programmers of each firm were engaged in. The situation
has been significantly changed with the advent of the 2. Overview
Windows operating system when the majority of the
developers switched to solving the problems of image The aim of object detection is to detect all instances of
processing itself. However, this has not yet led to the objects from a known class, such as people, cars or faces in
cardinal progress in solving typical tasks of recognizing an image. Generally, only a small number of instances of the
faces, car numbers, road signs, analyzing remote and object are present in the image, but there is a very large
medical images, etc. Each of these "eternal" problems is number of possible locations and scales at which they can
solved by trial and error by the efforts of numerous groups occur and that need to somehow be explored. Each detection
of the engineers and scientists. As modern technical of the image is reported with some form of pose
solutions are turn out to be excessively expensive, the task information. This is as simple as the location of the object, a
of automating the creation of the software tools for solving location and scale, or the extent of the object defined in
intellectual problems is formulated and intensively solved terms of a bounding box. In some other situations, the pose
abroad. In the field of image processing, the required tool kit information is more detailed and contains the parameters of
should be supporting the analysis and recognition of images a linear or non-linear transformation. For example for face
of previously unknown content and ensure the effective detection in a face detector may compute the locations of the
development of applications by ordinary programmers. Just eyes, nose and mouth, in addition to the bounding box of the
as the Windows toolkit supports the creation of interfaces face. An example of a bicycle detection in an image that
for solving various applied problems. specifies the locations of certain parts is shown in Figure 1.
The pose can also be defined by a three-dimensional
Object recognition is to describe a collection of related transformation specifying the location of the object relative
computer vision tasks that involve activities like identifying to the camera. Object detection systems always construct a
objects in digital photographs. Image classification involves model for an object class from a set of training examples. In
activities such as predicting the class of one object in an the case of a fixed rigid object in an image, only one
image. Object localization is refers to identifying the example may be needed, but more generally multiple
location of one or more objects in an image and drawing an training examples are necessary to capture certain aspects of
abounding box around their extent. Object detection does the class variability
work of combines these two tasks and localizes and

Volume 11 Issue 9, September 2022

www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 299
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942

Figure 1

Convolutional implementation of the sliding windows connected layers of the network into convolutional layers.
Before we discuss the implementation of the sliding window Fig.2 shows a simple convolutional network with two fully
using convents, let us analyze how we can convert the fully connected layers each of shape.

Figure 2: Simple convolution network

Object detection is an important task, yet challenging vision mode, include the brightness distortion and the chromaticity
task. It is a critical part of many applications such as image distortion which was used to distinguish shading background
search, image auto-annotation and scene understanding, from the ordinary background or moving foreground objects.
object tracking. Moving object tracking of video image The background and foreground subtraction method used the
sequences was one of the most important subjects in following approach. A pixel was modelled by a 4-tuple [Ei,
computer vision. It had already been applied in many si, ai, bi], where Ei-a vector with expected colour value, si-a
computer vision fields, such as smart video surveillance vector with the standard deviation of colour value, ai-the
(Arun Hampapur 2005), artificial intelligence, military variation of the brightness distortion and bi was the variation
guidance, safety detection and robot navigation, medical and of the chromaticity distortion of the ith pixel. In the next
biological application. In recent years, a number of step, the difference between the background image and the
successful single-object tracking system appeared, but in the current image was evaluated. Each pixel was finally
presence of several objects, object detection becomes classified into four categories: original background, shaded
difficult and when objects are fully or partially occluded, background or shadow, highlighted background and moving
they are obtruded from the human vision which further foreground object. Liyuan Li et al (2003), contributed a
increases the problem of detection. Decreasing illumination method for detecting foreground objects in non-stationary
and acquisition angle. The proposed MLP based object complex environments containing moving background
tracking system is made robust by an optimum selection of objects. A Bayes decision rule was used for classification of
unique features and also by implementing the Adaboost background and foreground changes based on inter-frame
strong classification method. colour co-occurrence statistics. An approach to store and fast
retrieve colour cooccurrence statistics was also established.
a) Background Subtraction In this method, foreground objects were detected in two
The background subtraction method by Horprasert et al steps. First, both the foreground and the background changes
(1999), was able to cope with local illumination changes, are extracted using background subtraction and temporal
such as shadows and highlights, even globe illumination differencing. The frequent background changes were then
changes. In this method, the background model was recognized using the Bayes decision rule based on the
statistically modelled on each pixel. Computational colour learned colour co-occurrence statistics. Both short-term and
Volume 11 Issue 9, September 2022
www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 300
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
long term strategies to learn the frequent background To circumvent the problem of selecting a huge number of
changes were used. An algorithm focused on obtaining the regions, Ross Girshick et al. proposed a method where we
stationary foreground regions as said by Álvaro Bayona et al use the selective search for extract just 2000 regions from
(2010), which was useful for applications like the detection the image and he called them region proposals. Therefore,
of abandoned/stolen objects and parked vehicles. This instead of trying to classify the huge number of regions, you
algorithm mainly used two steps. Firstly, asub-sampling can just work with 2000 regions. These 2000 region
scheme based on background subtraction techniques was proposals are generated by using the selective search
implemented to obtain stationary foreground regions. This algorithm which is written below.
detects foreground changes at different time instants in the
same pixel locations. This was done by using a Gaussian Selective Search:
distribution function. Secondly, some modifications were 1) Generate the initial sub-segmentation, we generate many
introduced on this base algorithm such as thresh holding the candidate regions
previously computed subtraction. The main purpose of this 2) Use the greedy algorithm to recursively combine similar
algorithm was reducing the amount of stationary foreground regions into larger ones
detected. 3) Use generated regions to produce the final candidate
region proposals
b) Existing Methods

Figure 3

These 2000 candid and fed into a convolutional neural within the region proposals, the algorithm the precision of
output. The CNN plays a role of feature extracted from the the bounding box. Have predicted the presence of a person
image and the extract object within that candidate region been cut in half. Therefore, the offset region proposal.

Figure 4: R-CNN

Problems with R-CNN a) Faster R-CNN

1) It still takes a huge amount of time to train the network
as you would have to classify 2000 region proposals per
image.
2) It cannot be implemented real time as it takes around 47
seconds for each test image.

Volume 11 Issue 9, September 2022

www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 301
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
image within the proposed region and predict the offset
values for the bounding boxes.

Figure 6: Comparison of test-time speed of object detection

algorithms

From the above graph, you can see that Faster R-CNN is
much faster than it‟s predecessors. Therefore, it can even be
used for real-time object detection.
Figure 5: Faster R-CNN b) YOLO — You Only Look Once
All the previous object detection algorithms have used
Both of the above algorithms (R-CNN & Fast Rxi-CNN) regions to localize the object within the image. The network
uses selective search to find out the region proposals. does not look at the complete image. Instead, parts of the
Selective search is the slow and time-consuming process image which has high probabilities of containing the object.
which affects the performance of the network. YOLO or You Only Look Once is an object detection
algorithm much is different from the region based
Similar to Fast R-CNN, the image is provided as an input to algorithms which seen above. In YOLO a single
a convolutional network which provides a convolutional convolutional network predicts the bounding boxes and the
feature map. Instead of using the selective search algorithm class probabilities for these boxes.
for the feature map to identify the region proposals, a
separate network is used to predict the region proposals. The
predicted the region which is proposals are then reshaped
using an RoI pooling layer which is used to classify the

Figure 7

YOLO works by taking an image and split it into an SxS class probability and offset values for the bounding box. The
grid, within each of the grid we take m bounding boxes. For bounding boxes have the class probability above a threshold
each of the bounding box, the network gives an output a value is selected and used to locate the object within the
Volume 11 Issue 9, September 2022
www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 302
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
image. 10) Matplotlib:
pip install matplotlib-command
YOLO is orders of magnitude faster (45 frames per second)
than any other object detection algorithms. The limitation of 11) H5py:
YOLO algorithm is that it struggles with the small objects pip install h5py
within the image, for example, it might have difficulties in
identifying a flock of birds. This is due to the spatial 12) Keras
constraints of the algorithm. pip install keras

System Requirement 13) ImageAI:

pip3 install imageai--upgrade
Install Python on your computer system
Install ImageAI and its dependencies like tensorflow, 14) Download the RetinaNet model file that will be used for
Numpy, OpenCV, etc. object detection using following link
1) Download the Object Detection model file (Retinanet) https://github.
2) Steps to be followed:- com/OlafenwaMoses/ImageAI/releases/download/1.0/resnet
3) Download and install Python version 3 from official 50_coco_best_v2.0.1. h5
Python Language website
https://python. org Copy the RetinaNet model file and the image you want to
detect to the folder that contains the python file.
4) Install the following dependencies via pip:
5) Tensorflow: InceptionV3:
pip install tensorflow-command Inception v3 is widely used as image recognition model that
has showed to obtain accuracy of greater than 78.1% on the
6) Numpy: ImageNet dataset. The model is the culmination of many
pip install numpy-command ideas developed by researchers over years. It is based on
“Rethinking the Inception Architecture Computer Vision”
7) SciPy: by Szegedy.
pip install scipy-command
The model is made of symmetric and asymmetric building
8) OpenCV: blocks, including convolutions, average pooling, max
pip install opencv-python-command pooling, concats, dropouts, and fully connected layers.
Batchnorm is used more throughout the model and applied
9) Pillow: to activation inputs. Loss is computed via Softmax.
pip install pillow -command
A high-level diagram of the model is shown below:

Figure 8

Figure 9: Transition layer

Volume 11 Issue 9, September 2022

www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 303
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
3. Results and Discussion

Figure 10: Before Detection

This is a sample image we feed to the algorithm and expect our algorithm to detect and identify objects in the image and label
them according to the class assigned to it.

Figure 11: After Detection

ImageAI provides many more features useful for 2) Custom Objects Detection: Using a provided
customization and production capable deployments for CustomObject class, you can tell the detection class to
object detection tasks. Some of the features supported are: report detections on one or a few number of unique
1) Adjusting Minimum Probability: By default, objects objects.
detected with a probability percentage of less than 50 3) Detection Speeds: You can reduce the time it takes to
will not be shown or reported. You can increase this detect an image by setting the speed of detection speed
value for high certainty cases or reduce the value for to “fast”, “faster” and “fastest”.
cases where all possible objects are needed to be 4) Input Types: You can specify and parse in file path to
detected. an image, Numpy array or file stream of an image as the

Volume 11 Issue 9, September 2022

www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 304
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
input image my sincere thanks to all of them.
5) Output Types: You can specify that the
detectObjectsFromImage function should return the I wish to express my deep sense of gratitude to my mentor,
image in the form of a file or Numpy array Dr. Nand Kumar for his able guidance and useful
Detection Speed: suggestions, which helped me in completing the project
Image AI now provides detection speeds for all object work, in time. He has been an immense source of
detection tasks. The detection speeds allow you to reduce the inspiration. Thanks a lot for his encouragement in carrying
time of detection at a rate between 20%-80%, and yet having out the project.
just slight changes but accurate detection results. Coupled
with lowering the minimum-percentage-probability My thanks and appreciations also go to my colleagues in
parameter, detections can match the normal speed and yet developing the project and people who have willingly helped
reduce detection time drastically. The available detection me out with their abilities: Khushbu Bharti, 20PGCS02
speeds are "normal" (default), "fast", "faster", "fastest"
and "flash". All you need to do is to state the speed mode References
you desire when loading the modelin the code.
detector. loadModel (detection_speed="fast") [1] Agarwal, S., Awan, A., and Roth, D. (2004). Learning
to detect objects in images via a sparse, part-based
Hiding/Showing Object Name and Probability: representation. IEEE Trans. Pattern Anal. Mach.
Image AI provides options to hide the name of objects Intell.26, 1475-1490. doi: 10.1109/TPAMI.2004.108
detected and / or the percentage probability from being [2] Alexe, B., Deselaers, T., and Ferrari, V. (2010). “What
shown on the saved/returned detected image. Using the is an object?, ” in ComputerVision and Pattern
detect Objects From Image () and detect Custom Objects Recognition (CVPR), 2010 IEEE Conference on (San
From Image () functions, the parameters display-object- Francisco, CA: IEEE), 73-80. doi:
name and display-percentage-probability can be set to True 10.1109/CVPR.2010.5540226
of False individually. [3] Aloimonos, J., Weiss, I., and Bandyopadhyay, A.
(1988). Active vision. Int. J. Comput. Vis.1, 333-356.
detections=detector.detectObjectsFromImage(input_image= doi: 10.1007/BF00133571
os.path.join(execution_path,"image3.jpg", [4] Andreopoulos, A., and Tsotsos, J. K. (2013).50 years
output_image_path=os.path.join(execution_path,"image3ne of object recognition: direc-tions forward. Comput.
w_nodetails.jpg"), minimum_percentage_probability=30, Vis. Image Underst.117, 827-891. doi: 10.1016/j.
display_percentage_probability=False, cviu.2013.04.005
display_object_name=False) [5] Azizpour, H., and Laptev, I. (2012). “Object detection
using strongly-superviseddeformable part models, ” in
4. Conclusion Computer Vision-ECCV 2012 (Florence: Springer),
836-849.
By using this thesis and based on experimental results we [6] Azzopardi, G., and Petkov, N. (2013). Trainable
are able to detect object more precisely and identify the cosfire filters for keypoint detectionand pattern
objects individually with exact location of an object in the recognition. IEEE Trans. Pattern Anal. Mach. Intell.35,
picture in x, y axis. This paper also provides experimental 490- 503. doi: 10.1109/TPAMI.2012.106
results on different methods for object detection and [7] Azzopardi, G., and Petkov, N. (2014). Ventral-stream-
identification and compares each method for their like shape representation: from pixel intensity values to
efficiencies. trainable object-selective cosfire models. Front.
Comput. Neurosci.8: 80. doi:
5. Future Enhancements 10.3389/fncom.2014.00080
[8] Benbouzid, D., Busa-Fekete, R., and Kegl, B. (2012).
For Night time visual tracking, night vision mode should be “Fast classification using sparsedecision dags, ” in
available as an inbuilt feature in the CCTV camera. Proceedings of the 29th International Conference on
MachineLearning (ICML-12), ICML „12, eds
To make the system fully automatic and also to overcome [9] J. Langford and J. Pineau (New York, NY:
the above limitations, in future, multi-view tracking can be Omnipress), 951-958.
implemented using multiple cameras. Multi view tracking [10] Bengio, Y. (2012). “Deep learning of representations
has the obvious advantage over single view tracking because for unsupervised and transferlearning, ” in ICML
of wide coverage range with different viewing angles for the Unsupervised and Transfer Learning, Volume 27 of
objects to be tracked. JMLRProceedings, eds I. Guyon, G. Dror,
[11] V. Lemaire, G. W. Taylor, and D. L. Silver (Bellevue:
In this thesis, an effort has been made to develop an JMLR. Org), 17-36.
algorithm to provide the base for future applications such as [12] Bourdev, L. D., Maji, S., Brox, T., and Malik, J.
listed below. (2010). “Detecting peopleusing mutually consistent
poselet activations, ” in Computer Vision - ECCV2010
Acknowledgement - 11th European Conference on Computer Vision,
incere efforts have been taken by me. However, it would not Heraklion, Crete, Greece, September 5-
have been possible without the kind support and help of [13] 11, 2010, Proceedings, Part VI, Volume 6316 of
many individuals and organizations. I would like to extend Lecture Notes inComputer Science, eds K. Daniilidis,
Volume 11 Issue 9, September 2022
www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 305
 International Journal of Science and Research (IJSR)
ISSN: 2319-7064
SJIF (2022): 7.942
P. Maragos, and N. Paragios (Heraklion: Springer),
168- 181.
[14] Bourdev, L. D., and Malik, J. (2009). “Poselets: body
part detectors trained using 3dhuman pose annotations,
” in IEEE 12th International Conference on
ComputerVision, ICCV 2009, Kyoto, Japan,
September 27 - October 4, 2009 (Kyoto: IEEE), 1365-
1372

Volume 11 Issue 9, September 2022

www.ijsr.net
Licensed Under Creative Commons Attribution CC BY
Paper ID: SR22804211151 DOI: 10.21275/SR22804211151 306