1OpenCV’s VideoCapture function

seize the real-time video feed from the camera that enabled us to access the camera and read
the video frames.

2Open Source Computer Vision Library.

obtaining the video frames, we preprocessed them using

resizing the frames to the necessary dimensions,

converting them into the required color space,

and normalizing the pixel values.

3 YOLO-NAS

After preprocessing, we input the frames into our model for object detection.

The model processed each frame and returned the coordinates of the bounding boxes for the
detected objects.

4 OpenCV’s drawing functions

to illustrate these bounding boxes on the original frames. We also labeled the boxes with the
classes of the detected objects.

5 Segment Anything Model (SAM)

to segment the objects in the frames..

6 Sort Algorithm

After obtaining the bounding boxes and segmentation masks, we had a lot of data points.

To manage this data effectively, we used the Sort Algorithm.

It helped us organize the bounding boxes and segmentation masks based on various criteria
such as their size, position, and class.

This sorting process made it easier for us to analyze the results and also improved the
efficiency of subsequent operations like tracking objects across frames.

7 OpenCV’s VideoWriter function

write the processed frames to an output video file.

This allowed us to create a video that visually represented the results of our object detection
and segmentation tasks.
reasons behind my choice of YOLO-NAS

Firstly, YOLO-NAS is the product of advanced Neural Architecture Search technology,

meticulously designed to address the limitations of previous YOLO models2. It introduces a
new quantization-friendly basic block3, which is designed to improve quantization performance
compared to its predecessors3. This new block allows YOLO-NAS to achieve higher accuracy
while maintaining efficiency3.

Secondly, YOLO-NAS significantly improves small object detection and localization accuracy3.
It also demonstrates superior performance in critical aspects such as small object detection,
localization accuracy, post-training quantization, and real-time edge-device applications1.

Thirdly, YOLO-NAS is open-source4 and is 10-20% faster than the pre-existing YOLO models4. It
uses a better architecture, AutoNAC4, which sets a new record in object detection, providing
the best accuracy and latency tradeoff performance4.

Now, let’s talk about the accuracy of YOLO-NAS. The model, when converted to its INT8
quantized version, experiences a minimal precision drop, a significant improvement over other
models2. In terms of pure numbers, YOLO-NAS is ~0.5 mAP point more accurate and 10-20%
faster than equivalent variants of YOLOv8 and YOLOv756.
role of OpenCV
OpenCV, or Open Source Computer Vision, is a powerful library that we utilized
extensively in our project. It provided us with tools to process images videos, which
are crucial in the field of object detection.
One of the primary uses of OpenCV in our project was in preprocessing the input
images. Before feeding the images into our YOLO-NAS model, we used OpenCV to
resize the images to the required dimensions, and normalize the pixel values. This
ensured that our model received inputs in a consistent and standardized format,
thereby improving its performance.
We also used OpenCV for drawing bounding boxes and labels on our output images.
Once our YOLO-NAS model detected objects in an image, we used OpenCV’s drawing
functions to visually represent these detections. This involved drawing bounding
boxes around the detected objects and labeling them with their respective classes.
This visual representation made it easier for us to interpret the model’s output and
verify its accuracy.
In addition, OpenCV was instrumental in handling real-time video streams. We used
its VideoCapture and VideoWriter classes to read input from a video file or a camera
in real-time, run our YOLO-NAS model on each frame, and then write the output
frames (with the detections drawn on them) to an output video file. This allowed us
to use our model for real-time object detection tasks.
Lastly, OpenCV’s vast array of image processing functions were invaluable
in augmenting our training data. We used functions like rotation, translation, scaling,
and flipping to artificially increase the size of our training dataset and introduce
more variability into it. This helped improve the robustness of our model and its
ability to generalize to new, unseen data.
Segment Anything Model (SAM)**,

segmentation model developed by Meta's FAIR (Fundamental AI Research) lab¹.

Firstly, SAM's **unmatched versatility**⁵ and **zero-shot inference capabilities**¹ were

instrumental in our project. Unlike traditional models that require extensive training on specific
tasks, SAM can accurately segment images without prior specific training¹. This means that
SAM could handle a wide array of objects across different images and videos in our project,
regardless of its training data⁵.

Secondly, SAM's **high accuracy**³ in segmenting objects in images was a key advantage. With
over 1 billion masks on 11M licensed and privacy-respecting images, SAM’s zero-shot
performance is often competitive with or even superior to prior fully supervised results². This
ensured that our project benefited from precise and reliable segmentation results.

In conclusion, the Segment Anything Model (SAM) has been an invaluable tool in our project. Its
versatility, high accuracy, ease of use, and real-time interaction capabilities have significantly
enhanced the performance and efficiency of our project.
Sort Algorithm

Firstly, the Sort Algorithm was used in organizing our data.. However, unorganized data
can lead to inefficiency and inaccuracies. By using the Sort Algorithm, we were able to
arrange our data in a specific order, be it numerical, alphabetical, or based on any other
relevant criteria. This made our data more manageable and accessible.

Thirdly, the Sort Algorithm was instrumental in identifying patterns and making
predictions. Once our data was sorted, it was easier to observe trends and patterns. This was
particularly useful in the data analysis phase of our project, where we needed to make
informed decisions based on our data.

Lastly, the Sort Algorithm helped in improving the user experience. For instance, if our
project involved displaying information to the user, sorted data allowed us to present this
information in a more structured and understandable manner.

In conclusion, the Sort Algorithm was an invaluable tool in our project. Its ability to organize
data, optimize search operations, identify patterns, and improve user experience significantly
enhanced the efficiency and effectiveness of our project.
a responsive alerting mechanism

In our project, we developed a responsive alerting mechanism for timely threat notification.
This process began with the detection of potential threats in the video feed using our YOLO-
NAS model. When a threat was identified, the model generated a bounding box around it and
classified it.

Following the detection, the Segment Anything Model (SAM) was used to segment the
detected threat from the rest of the image. This segmentation provided us with more granular
information about the threat, such as its exact shape and size.

Once the threat was detected and segmented, we evaluated it based on various criteria such as
its size, position, and class. This evaluation helped us to determine the severity of the threat.

If the threat issevere enough, an alert was generated. This alert contained information about
the threat, such as its class, position, and size. The generated alert was then sent to the
relevant parties through an appropriate notification channel. This could be an email, a push
notification, or any other form of communication that was suitable for the situation.