Scribd
Upload
5 views12 pages

Final Report Submission For Group 14: Course: Internet of Things (EEE F411)

Uploaded by

f20220630
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
5 views12 pages

Final Report Submission For Group 14: Course: Internet of Things (EEE F411)

Uploaded by

f20220630
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 12

FINAL REPORT SUBMISSION

FOR GROUP 14

Course: Internet of Things (EEE F411)

Members:
Shashwata Ghosh: 2022B5A31036H
Divyansh Rajawat: 2022B5A31161H
Vishwajith Sunkerla: 2021A3PS2581H
Nilesh Gupta: 2021A3PS1699H
Rashi Singh: 2021A3PS0807H
Summary of what we did up till
mid-sem:

The papers we reviewed were based on the use of edge


computing in the field of autonomous cars. The 5 papers
are listed below:-
● A camera radar fusion model based on edge
computing
● Task offloading of deep learning services for
autonomous driving in mobile edging computing
● Multi access edge computing: the driver behind the
wheel of 5g-connected cars
● Dynamic obstacle avoidance model of autonomous
driving with attention mechanism and temporal
residual block
● Use of machine learning in detecting network security
of edge computing system

Paper we worked on:

We chose the paper titled: “Dynamic obstacle


avoidance model of autonomous driving with
attention mechanism and temporal residual block” to
work on for the rest of our project.
What we wanted to improve on was the accuracy of
identifying incoming obstacles in the path of an
autonomous vehicle to improve safety.

What is our problem statement?

● Dynamic obstacle avoidance in autonomous driving faces


challenges in complex environments, including missed
detections and unstable tracking, particularly for high-speed
or occluded targets.
● Current deep learning methods struggle with high accuracy
due to tracker performance limitations.
● The paper proposes a novel YTCN model using Temporal
Convolutions to incorporate time-series data, improving
obstacle detection and tracking.
● Enhanced feature extraction is achieved through the
SPPFCSPC structure.
● Testing shows the YTCN model significantly reduces target
loss and increases tracking accuracy.

What are we doing?

Our main focus was to find a good model for object


detection and implementation of the model employing
YoloV8, temporal residual block and SPPFCSPC
structure.
This involved 2 main steps:
1. Model building(data training)
2. Dataset test

Model Building- pipeline configuration:


Data Test:
One of the pictures we tested the
model on:
Second solution to import a better
model than YoloV8 and build upon
the model:
This is a more advanced model which uses RCNN on top
of the technical YoloV8 so this would be an alternative
solution to the normal object detection methods

Drawbacks of the proposed model:


The main issue for the implementation of this paper was
the level of complexity and hardware requirements for
building the model.
While the BDD100K dataset is extensive and diverse, the
paper's reliance on it may limit the evaluation of the
model's generalizability to other real-world datasets or
non-urban environments.
Executing such models on every setup is time consuming
and resource draining.
We tried training and testing it with a few photos to
determine our accuracy and nobility of our project.

What did we learn from this


project:
Implementing YOLO NAS and YOLOV8 has taught us the
following:-
1. Object detection basics like boundary boxes and
difference between one stage and two stage detectors
2. We learnt about the attention mechanisms, feature
pyramids and backbone networks used in both
models.
3. We learnt about applying custom datasets.
4. We learnt about the basics of machine learning and
terms like feature extraction.

Conclusion
The paper proposed was implemented up till
SPPSFCSPC structure but the model prepared by us
wasn’t giving accurate solutions.
We looked into another solution on the internet, and found
a modified model of the YoloV8: YOLO NAS
It has a better mAP and lower latency.
YOLO-NAS's architecture employs quantization-aware
blocks and selective quantization for optimized
performance. When converted to its INT8 quantized
version, YOLO-NAS experiences a smaller precision drop
(0.51, 0.65, and 0.45 points of mAP for S, M, and L
variants) compared to other models that lose 1-2 mAP
points during quantization. These techniques culminate in
innovative architecture with superior object detection
capabilities and top-notch performance.

Contribution of every person:

1. Shashwata Ghosh: helped with the implementation


of YOLOV5, comparing models and preparing the
presentation and PPT.
2. Divyansh Rajawat: helped with the implementation
of YOLOV5, comparing models and preparing the
presentation and PPT.
3. Viswajith Sunkerla: helped with implementation of
YOLOV8 and YOLO NAS.
4. Nilesh Gupta: helped with the implementation of
yolov8
5. Rashi Singh: helped with testing.

You might also like