Task-1 --> Drone Object Detection

Let's understand how we collected the dataset and trained the CNN->(Convulation Neural Network) model on that dataset and predict the drones or small drones present in the images and videos.
We can also use the saved model after the training to predict the drones in any any type of input either live video streams from cameras, or recorded videos.

Step 1: Data Collection and Preprocessing

Collect a dataset containing images or video frames with small drones as the target objects. Annotating these datasets with bounding boxes indicating the location of drones. Then preprocessing the dataset by resizing images to a fixed size, normalizing pixel values, and augmenting data to increase dataset diversity (e.g., rotation, flipping, scaling).
So the dataset that I collected was from the Roboflow open source dataset uploaded by some user which contains around 13800 images of Training and 3000 images Validation with the information of there respective boundary boxes.
I also collected a separate test data which contains around 163 images to see the performance of the fine tuned model.

Now Design a CNN architecture suitable for small object detection and tracking. We can base our design on popular architectures like YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), or Faster R-CNN (Region-based Convolutional Neural Networks).
After that we have to ensure that the architecture is lightweight and optimized for real-time performance on devices like drones. This may involve reducing the number of layers, using smaller filter sizes, or employing depthwise separable convolutions.
I used YOLO model which is based on CNN and the version that I used is v8.0.196.
This is the YOLOV8 architecture, you can see that this is an Anchor free moedel which means it directly predicts the centre of the image instead of the offsets called anchor box.

And the model that I was traing consist of 295 layers and around 25856899 parameters , and working at 79.1 GFLOPS.

I trained the model for 10 epochs(you can train for more number of epochs like 50 or 100 with a larger Batch size, I used a image batch of 512 images, we can also use 800 or more if you have the larger and powerful GPUs.
Optimization is Adamw optimizer and learning rate = 0.002.
After the fine tuning of the model the model is saved(best.pt) and we can use that to predict on images and video streams.

Evaluate the trained model on a separate test set to assess its performance in terms of detection accuracy, tracking stability, and real-time inference speed.
Then I computed metrics like precision, recall, F1-score, and Confusion matric to see the how the model predicted on the validation and test images.
Then I analyze the results and iteratively refine the model architecture and training process based on the observed performance.
access the loss function in the working/runs/detect/train/result.png --> see it is for 10 epochs only and the we can increase the epochs to reduce the errors.
access the confusion matrix in working/runs/detect/train/confusion_matrix.png

After that I uploaded 2 video from the youtube to see how the model performs.
You can access the output from the model it is save in the dir working/runs/detect/predict4/ and working/runs/detect/predict2/