Detection and Classification
using Deep Models
Presented
By
Komal Vishnu Venkata Reddy
Outline
• Introduction
• Challenges
• Research Gaps
• Research Works
• Research Ongoing
• Conclusion
Introduction
Automatic vehicle detection and classification (AVDC)
systems have become important for real-time traffic
monitoring and management.
AVDC requires collection of suitable data with real time
traffic information, and automatic vehicle classification and
detection methodologies.
Fig 1. Vehicle Detection
Research Challenges
Image and video
datasets have its
own challenges.
Making models
that are robust
enough to use for
inputs taken in
different weather
conditions.
Models dealing
with data in
daylight conditions
might not work for
data taken in night
conditions.
Due to congested
traffic, a single
frame has multiple
vehicles.
Research Gaps
• Many datasets used over the years give almost 100% accuracy.
• Huge samples are needed to train deep learning based models.
• Many datasets are not accurately processed and annotated .
• In Indian sub-continent road conditions, vehicle types, traffic
scenarios are dissimilar than those found in developed
countries.
• Hence, most of the research articles available in the literature
may not be suitable for such cases.
• Some datasets have very few classes.
Last Decade in
Vehicle Detection
and
Classification: A
Comprehensive
Survey
• The current survey encompasses all key research papers published
on AVD between the year 2010 and 2020.
• It gives a general overview of the several approaches applied for
both localizing as well as recognizing different vehicle classes
using both machine learning and deep learning approaches.
• It includes the most widely used AVD datasets (comprising of
video as well as still-image datasets) used for both vehicle
localization as well as classification problems.
• It also presents a critical review of the deep learning based methods, which is the
current trend, used for AVD problem.
• Finally, it analyses different prospects of future research achievements in this
field
• Published in Archives of Computational Methods in Engineering -Springer
Two decades of
vehicle make and
model
recognition –
Survey,
challenges and
future directions
• To the best of our knowledge, this paper discusses most
of the research papers on VMMR published between the
year 2004 and 2023.
• We have identified as well as provided a general
overview of various methodologies for VMMR based on
both machine learning and deep learning models
supported by some in-depth discussion.
• We have provided information regarding several vehicle
related datasets used for VMMR task.
• Finally, we have presented various research gaps, and
their potential solutions for the said field’s research
advancements in the future.
• Published in Journal of King Saud University -
Computer and Information Sciences - Sciencedirect -
Elsevier
Current Datasets and Their Inherent Challenges for Automatic
Vehicle Classification
• Also, we have given a comparative study of the
different types of datasets used for
classification along with their pros and cons.
• This study presents a comprehensive survey of
the datasets available for AVC and vehicle
model and make recognition (VMMR)
published in the last 10 years highlighting their
inherent challenges.
• Published in Machine Learning for Cyber
Physical System: Advances and Challenges-
Springer
Performance Comparison of Various
YOLO Models for Vehicle Detection:
An Experimental Study
• In this paper, we focus on three major object
detection algorithms under the YOLO family,
namely YOLOv5, YOLOv7, and YOLOv8 for the
purpose of vehicle detection,
• Discuss the architectural differences of these
variants.
• Performance comparison of these models, and
in doing so, we use two recently introduced AVD
datasets developed for the Indian subcontinent,
namely JUVDsi v1 and IRUVD.
• Published in Proceedings of Data Analytics and
Management- ICDAM 2023, Volume 3 Springer
JUVDsi v1: developing and benchmarking a new
still
image database in Indian scenario for automatic
vehicle detection
• The image database is properly annotated to measure
the performance of any algorithm developed for
automatic localization and classification of vehicles in
an unconstrained environment.
• Different complexities are added to the images to make
it challenging as well as realistic.
• Nine different classes of vehicles are presented in this
database – very few of the existing databases have
considered such a variety of vehicle classes.
• Three models namely, You Only Look Once (YOLO) ,
Region-Based Convolutional Neural Networks (R-CNN) ,
and Region-Based Fully Convolutional Networks (RFCN)
are used. Finally an ensemble method, called Weighted
Boxes Fusion (WBF) is implemented.
• Published in Multimedia Tools and Applications,
Springer
JUIVCDv1: development of a still-image
based dataset
for indian vehicle classification
• This dataset offers a realistic image representation of the
traffic situation in India, which isbvery different from that of
other developed countries. Vehicle images captured in
variousbscenarios are considered. A total number of 6335
vehicle images can be found in this dataset..
• Researchers may take this dataset to evaluate the
effectiveness of their methods for autonomous vehicle
localization and categorization.
• The vehicle images in the collection are taken in different
weather conditions. Therefore, the model is resilient
enough to handle data collected in a variety of
meteorological scenarios.
• We have benchmarked this dataset using an MVE classifier
combination approach which achieves 95% accuracy.
• Published in Multimedia Tools and Applications, Springer
XMR_Net: A Deep Model for Vehicle Make
and Model Recognition using Still-images
• In this work, we have proposed an ensemble of
attention-aided three deep CNN models, called
XMR_Net, for VMMR.
• In this paper, initially, we have used five standard
convolutional neural network (CNN) models, namely
Inceptionv3, Xception, InceptionResNetv2,
MobileNetV2, and ResNet152v2 for VMMR.
• We have also used an attention mechanism to these
models. To increase accuracy of the overall model
• we have chosen three best base learners from these
five CNN models, and formed an ensemble model. The
final model is called XMR_Net
SimSANet: A Simple Sequential Attention aided Deep Neural Network
for Vehicle Make and Model Recognition
• We present Simple Sequential Attention Network (SimSANet),
a multikernel-based sequential attention-based model, which
efficiently extracts the most discriminative information by
combining both global and local features.
• It also offers significant advantages in speed, effectiveness,
and efficiency, requiring far fewer parameters compared to
existing models.
• To demonstrate the significance of each layer in the
architecture of the suggested model, along with the
recommended values for the hyperparameters, we perform a
statistical analysis and ablation studies. We utilize the Grad-
CAM to show the efficacy of the proposed model.
• We conduct extensive experiments on multiple public VMMR
benchmark datasets to ensure the effectiveness of the
proposed model.
A Feature Fusion based Custom Deep Learning
Model for Vehicle Make and Model Recognition
• Deep Feature Fusion: It fuses two feature maps
that are extracted from input images in two parallel
paths containing two different base models.
• Modified CBAM: Taking inspiration from the
original CBAM attention [14], a modified attention
mechanism is used, which extracts channel and
spatial information in parallel and combines them
with input feature maps.
• Generalized Approach: Our model is evaluated on
two VMMR datasets, namely Stanford Cars and
Comp-CarsSV, and achieved 93.51% and 99.03%
test accuracies, respectively.
Research
Ongoing
• Developing vehicle detection
dataset for adverse weather
condition.
• Developing a deep learning based
vehicle detection lightweight
model.
RESULTS
• . High-Accuracy Classification via Transfer Learning: The core of
the project is a deep learning model that classifies vehicle images.
Instead of building a model from scratch, we used transfer
learning with Google's MobileNetV2, a state-of-the-art Convolutional
Neural Network (CNN). This allowed us to leverage a model pre-trained
on millions of images, leading to faster training and higher accuracy on
our specific task of identifying bikes, buses, cars, and trucks.
• Robust Web Application with Flask: The trained model was deployed
in a web application built with the Python Flask framework. The
backend handles image uploads, preprocesses them to the required
224x224 pixel size, and uses the saved vehicle_model.h5 file to make
predictions. This creates a practical and user-friendly interface for the AI
model.
Result
• An image recognition model
identifies a truck by detecting its
most defining characteristics,
primarily its large, two-part
structure consisting of a distinct
cab and a separate, long cargo
area. It also recognizes key local
features like high ground
clearance, a large vertical front
grille, and multiple sets of wheels,
which differentiate it from smaller
cars or single-body buses.
Conclusion
MAIN AIM IS TO MAKE AN EFFICIENT AVDC
SYSTEM THAT CAN BE USED TO SOLVE THE
AMBIGUITY ISSUES DUE TO OVERLAPPING
VEHICLES.
SYSTEMS SHOULD BE WORKING EFFICIENTLY ON
IMAGES AND VIDEOS TAKEN IN DIFFERENT
WEATHER CONDITIONS.
DATASETS WILL BE DEVELOPED FOCUSING ON
INDIAN TRAFFIC AND ROAD CONDITIONS.
Thank You

vehicle recognisation using deep learning cnn.pptx

  • 1.
    Detection and Classification usingDeep Models Presented By Komal Vishnu Venkata Reddy
  • 2.
    Outline • Introduction • Challenges •Research Gaps • Research Works • Research Ongoing • Conclusion
  • 3.
    Introduction Automatic vehicle detectionand classification (AVDC) systems have become important for real-time traffic monitoring and management. AVDC requires collection of suitable data with real time traffic information, and automatic vehicle classification and detection methodologies. Fig 1. Vehicle Detection
  • 4.
    Research Challenges Image andvideo datasets have its own challenges. Making models that are robust enough to use for inputs taken in different weather conditions. Models dealing with data in daylight conditions might not work for data taken in night conditions. Due to congested traffic, a single frame has multiple vehicles.
  • 5.
    Research Gaps • Manydatasets used over the years give almost 100% accuracy. • Huge samples are needed to train deep learning based models. • Many datasets are not accurately processed and annotated . • In Indian sub-continent road conditions, vehicle types, traffic scenarios are dissimilar than those found in developed countries. • Hence, most of the research articles available in the literature may not be suitable for such cases. • Some datasets have very few classes.
  • 6.
    Last Decade in VehicleDetection and Classification: A Comprehensive Survey • The current survey encompasses all key research papers published on AVD between the year 2010 and 2020. • It gives a general overview of the several approaches applied for both localizing as well as recognizing different vehicle classes using both machine learning and deep learning approaches. • It includes the most widely used AVD datasets (comprising of video as well as still-image datasets) used for both vehicle localization as well as classification problems. • It also presents a critical review of the deep learning based methods, which is the current trend, used for AVD problem. • Finally, it analyses different prospects of future research achievements in this field • Published in Archives of Computational Methods in Engineering -Springer
  • 7.
    Two decades of vehiclemake and model recognition – Survey, challenges and future directions • To the best of our knowledge, this paper discusses most of the research papers on VMMR published between the year 2004 and 2023. • We have identified as well as provided a general overview of various methodologies for VMMR based on both machine learning and deep learning models supported by some in-depth discussion. • We have provided information regarding several vehicle related datasets used for VMMR task. • Finally, we have presented various research gaps, and their potential solutions for the said field’s research advancements in the future. • Published in Journal of King Saud University - Computer and Information Sciences - Sciencedirect - Elsevier
  • 8.
    Current Datasets andTheir Inherent Challenges for Automatic Vehicle Classification • Also, we have given a comparative study of the different types of datasets used for classification along with their pros and cons. • This study presents a comprehensive survey of the datasets available for AVC and vehicle model and make recognition (VMMR) published in the last 10 years highlighting their inherent challenges. • Published in Machine Learning for Cyber Physical System: Advances and Challenges- Springer
  • 9.
    Performance Comparison ofVarious YOLO Models for Vehicle Detection: An Experimental Study • In this paper, we focus on three major object detection algorithms under the YOLO family, namely YOLOv5, YOLOv7, and YOLOv8 for the purpose of vehicle detection, • Discuss the architectural differences of these variants. • Performance comparison of these models, and in doing so, we use two recently introduced AVD datasets developed for the Indian subcontinent, namely JUVDsi v1 and IRUVD. • Published in Proceedings of Data Analytics and Management- ICDAM 2023, Volume 3 Springer
  • 10.
    JUVDsi v1: developingand benchmarking a new still image database in Indian scenario for automatic vehicle detection • The image database is properly annotated to measure the performance of any algorithm developed for automatic localization and classification of vehicles in an unconstrained environment. • Different complexities are added to the images to make it challenging as well as realistic. • Nine different classes of vehicles are presented in this database – very few of the existing databases have considered such a variety of vehicle classes. • Three models namely, You Only Look Once (YOLO) , Region-Based Convolutional Neural Networks (R-CNN) , and Region-Based Fully Convolutional Networks (RFCN) are used. Finally an ensemble method, called Weighted Boxes Fusion (WBF) is implemented. • Published in Multimedia Tools and Applications, Springer
  • 11.
    JUIVCDv1: development ofa still-image based dataset for indian vehicle classification • This dataset offers a realistic image representation of the traffic situation in India, which isbvery different from that of other developed countries. Vehicle images captured in variousbscenarios are considered. A total number of 6335 vehicle images can be found in this dataset.. • Researchers may take this dataset to evaluate the effectiveness of their methods for autonomous vehicle localization and categorization. • The vehicle images in the collection are taken in different weather conditions. Therefore, the model is resilient enough to handle data collected in a variety of meteorological scenarios. • We have benchmarked this dataset using an MVE classifier combination approach which achieves 95% accuracy. • Published in Multimedia Tools and Applications, Springer
  • 12.
    XMR_Net: A DeepModel for Vehicle Make and Model Recognition using Still-images • In this work, we have proposed an ensemble of attention-aided three deep CNN models, called XMR_Net, for VMMR. • In this paper, initially, we have used five standard convolutional neural network (CNN) models, namely Inceptionv3, Xception, InceptionResNetv2, MobileNetV2, and ResNet152v2 for VMMR. • We have also used an attention mechanism to these models. To increase accuracy of the overall model • we have chosen three best base learners from these five CNN models, and formed an ensemble model. The final model is called XMR_Net
  • 13.
    SimSANet: A SimpleSequential Attention aided Deep Neural Network for Vehicle Make and Model Recognition • We present Simple Sequential Attention Network (SimSANet), a multikernel-based sequential attention-based model, which efficiently extracts the most discriminative information by combining both global and local features. • It also offers significant advantages in speed, effectiveness, and efficiency, requiring far fewer parameters compared to existing models. • To demonstrate the significance of each layer in the architecture of the suggested model, along with the recommended values for the hyperparameters, we perform a statistical analysis and ablation studies. We utilize the Grad- CAM to show the efficacy of the proposed model. • We conduct extensive experiments on multiple public VMMR benchmark datasets to ensure the effectiveness of the proposed model.
  • 14.
    A Feature Fusionbased Custom Deep Learning Model for Vehicle Make and Model Recognition • Deep Feature Fusion: It fuses two feature maps that are extracted from input images in two parallel paths containing two different base models. • Modified CBAM: Taking inspiration from the original CBAM attention [14], a modified attention mechanism is used, which extracts channel and spatial information in parallel and combines them with input feature maps. • Generalized Approach: Our model is evaluated on two VMMR datasets, namely Stanford Cars and Comp-CarsSV, and achieved 93.51% and 99.03% test accuracies, respectively.
  • 15.
    Research Ongoing • Developing vehicledetection dataset for adverse weather condition. • Developing a deep learning based vehicle detection lightweight model.
  • 16.
    RESULTS • . High-AccuracyClassification via Transfer Learning: The core of the project is a deep learning model that classifies vehicle images. Instead of building a model from scratch, we used transfer learning with Google's MobileNetV2, a state-of-the-art Convolutional Neural Network (CNN). This allowed us to leverage a model pre-trained on millions of images, leading to faster training and higher accuracy on our specific task of identifying bikes, buses, cars, and trucks. • Robust Web Application with Flask: The trained model was deployed in a web application built with the Python Flask framework. The backend handles image uploads, preprocesses them to the required 224x224 pixel size, and uses the saved vehicle_model.h5 file to make predictions. This creates a practical and user-friendly interface for the AI model.
  • 17.
    Result • An imagerecognition model identifies a truck by detecting its most defining characteristics, primarily its large, two-part structure consisting of a distinct cab and a separate, long cargo area. It also recognizes key local features like high ground clearance, a large vertical front grille, and multiple sets of wheels, which differentiate it from smaller cars or single-body buses.
  • 18.
    Conclusion MAIN AIM ISTO MAKE AN EFFICIENT AVDC SYSTEM THAT CAN BE USED TO SOLVE THE AMBIGUITY ISSUES DUE TO OVERLAPPING VEHICLES. SYSTEMS SHOULD BE WORKING EFFICIENTLY ON IMAGES AND VIDEOS TAKEN IN DIFFERENT WEATHER CONDITIONS. DATASETS WILL BE DEVELOPED FOCUSING ON INDIAN TRAFFIC AND ROAD CONDITIONS.
  • 19.