Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 12

Challenge in India

Ideas around Intelligent Transportation Systems have been in


the pipeline for many cities across India, primarily for the metro
cities. However, each endeavour mostly utilized only certain
aspects of ITS, not implementing it in its entirety. While some
projects were an intermediate success, many failed to create a
significant impact. Lack of proper implementation, regulation
and compliance with the law were some of the major
contributing factors to these versions of ITS not working out.
Intelligent Transport Systems use the latest tech trends such
as Computer Vision, Deep Learning, edge computing (taught in
class), IoT(taught in class), etc., to power such solutions.
Challenge in ITS in India
There is a company that develops intelligent traffic systems
(ITS) for government, police, and traffic departments. They
work as a substantial computer vision expertise ( from
discovery to implementation) for projects.

They are currently working on an outdoor Intelligent Transport


System aimed to track whether there are no traffic rules
offenses. Cameras are located on specific road parts that detect
the car, speed, color, and size. Their task is to develop models
that will detect a specific anomaly based on the image received
from the cameras.
As that is an outdoor system, they come across a few
challenges that we needed to solve:

•Different weather conditions (rain, snow, sun, clouds)


•Different light conditions (day, night, evening)
•Windshield reflections (hard to see passengers, drivers)
•Left, right-sided cars (hard to detect a fastened seat belt)
•Data collection:
All data collection for training models is performed on the
company’s side. The data (snapshots with different time
intervals) is retrieved from Jetson TX2(AI computing) devices
(with connected cameras) and sent directly to AWS S3, where it
is stored, and from which we upload the images for further
annotation and analysis.
•Data annotation:
This stage can be both manual and automated and is required
for further system training.

It typically consists of several phases:


1.With the help of the CVAT tool, we perform manual
annotation. The tool also allows uploading automatic
annotation using models we already trained on smaller
datasets. For instance, we annotate images according to the
windshield/person detector, seatbelt/offense classifier. In
simple terms, we annotate if there are any offenses on the
images: eating, drinking, talking over the phone, excessive head
rotation, hands on the wheel, or not.
2.We manually annotate or fix automatic annotations. 
3.We download correct annotations in COCO format from CVAT.
4.We use data analytics for Intelligent Transportation Systems
to study the annotated data and use it for retraining models
and making the results more accurate.
•Data analysis:
Data undergoes steps such as error rectification, data cleaning,
data synthesis, and adaptive logical analysis during this stage.
Inconsistencies in data are identified with specific software and
rectified. After that, data is pooled for analysis and analyzed
further to train models, predict traffic scenarios, and detect
anomalies.
To ensure that the model is trained on accurate data, we
analyze annotations using the Pandas tool. It is crucial to
understand how many images we have collected, and how
often specific objects (i.e., windshields, drivers, passengers are
presented on the images, what the minimum/maximum size of
these objects is. How many images we have with different
cases(e.g., with a seatbelt fastened or not fastened, and
whether the dataset with these cases is balanced).
•Image enhancement:
Also, we helped our client with image enhancement, as we
work with outdoor cameras during different weather conditions
and dayparts. So, we needed to reduce reflections, brightness,
darkness, and other factors that affect image quality.

•Training, validating, and testing models:


1.Collected, annotated, and analyzed data is used as the
training dataset. Retraining models as the training dataset
grows to increase the accuracy of models.
2.Evaluating the accuracy of models on the testing dataset to
see if they need to be retrained to provide more accurate
results.

We require strong team of expertise in this


TECH TRENDS THAT POWER INTELLIGENT TRANSPORT SYSTEMS

•Deep Learning 
Deep learning is a type of Machine Learning that is used for working with unstructured
data, such as images and video. That’s why it is widely used in Intelligent transport systems
for image detection (e.g., windshield, person detection) and classification(e.g., offense,
seat belt) in Intelligent Transport Systems.

Here are some of the critical Deep Learning tools that we use in Intelligent transport
systems:
• Tensorflow Object Detection API and Transfer Learning Toolkit: for object detection
(cars, pedestrians, traffic lights, etc.);
• Deep Stream: allows the implementation of video surveillance systems on the road,
but it is limited only to specific deep learning models and camera models. So it may not
be a good fit in some cases.
• Computer Vision
• The Computer Vision tool, OpenCV, is used for image enhancement and preprocessing
(resizing, cropping, normalizing, removing distortions). Also, Computer Vision is needed
for recognizing different weather conditions. 
• Internet of Things
• IoT is used for different purposes in Intelligent Transport Systems. For instance, an AI-
powered embedded system Jetson TX2 is a single board computer to which a camera is
connected. It is used for processing the data on-premise to ensure better security of
personal data. What’s more, it doesn't require transferring large volumes of data and
therefore can do with 4G. Only the processed results and alerts are sent to the cloud.
That is why it is a perfect fit for Intelligent Transport Systems.
BEST PRACTICES TO IMPLEMENT SUCCESSFUL INTELLIGENT TRANSPORT SYSTEMS
1.Define goals and have a clear understanding of business KPIs. 
A Discovery Stage to help you understand what to start with, the risks and challenges, how to overcome them, what tech stack and architecture
are the best fit for your specific case (taking into account the types of cameras and hardware you are currently using). As a result, you mitigate
potential risks and get all the deliverables to kick off the project successfully.
2.When training models, ML engineers, and Data scientists need to have a clear understanding of the data and challenges associated with
retrieving clean, valid, and complete data.
ML engineers and Data scientists need to ensure the dataset for training is balanced (each class should be present in the dataset in equal parts).
For instance, to ensure a model detects correctly if the seatbelt is fastened or not, you need to have a balanced dataset to train it on. It means
you need an equal number of images when: 
3.a seatbelt is fastened;
4.a seatbelt is not fastened. 
It allows training models that will be more accurate.
3.Prepare a nice and clean dataset and use transfer learning. It is critical to focus on data quality, not quantity. 
You can start training backbone models (available from Google, Amazon) without collecting thousands of data/images by yourself as they have
already been trained on billions of data. And they can already detect and identify colors, patterns, lines, etc.  All you need is to customize them
and start training on custom data (even beginning with 100 images) and then retrain them as the pool of your images increases.
4.Prototype fast and increase the training dataset step by step (for various conditions, in different locations), re-train models as the training
dataset is growing, compare the accuracy of newly trained models to the previous ones, and retrain as much as you need to reach the required
level of accuracy.
5.Use efficient models to fit in memory and performance limits and use multitask learning to save memory and improve performance. It means
that you can use a single backbone model as a feature extractor but with different tails for tasks (for example: do seatbelt/offense classification
with a single model but different tails).
6.Use the automation of pipelines and implement MLOps. That will allow you to implement continuously repeatable cycles of training, testing,
deploying, monitoring, and operating the ML models, do more experiments, and increase the accuracy of models more time-and-cost efficiently. 
7.Build a PoC with baseline accuracy. Then, by using incremental learning, improve models gradually to reach a baseline accuracy of 100%. 
How to ensure security in Intelligent Transport Systems

•The team works remotely on AWS machines without an


Internet connection. The only access on machines is to data (S3
Storage) and code (ssh-based connection to bitbucket repo).
•All packages are prepared and installed offline after the IT
Security Team transfers data to the machine. 
•Use Docker containers to isolate different environments.

You might also like