DylanC7/CS370-HW3

I used the dataset "Drone Computer Vision Model" from roboflow https://universe.roboflow.com/project-986i8/drone-uskpc I felt this model had the best attributes to complete the assignment, it was trained on a variety of pictures and all the metrics like Recall Map and percision seemed very accurate.

For my detector i used Yolov8n which after doing some research seemed to be the best fit for tracking smaller object. it also good for faster inference. I fined tuned the model using 20 epochs, imgsz=640, batch=16, GPU T4 (google colab GPU). I wanted to keep the resolution but i chose a lower epochs size because the original dataset was trained on a vast amount of images (also it would have been tough to run larger).

For the Kalman filter design i did my state vector using the center position of the drone / box (cx,cy) and then i did the velocity it is traveling at. This allows the model to help estimate where it would be at a certain osition because I did the sampling every 5 FPS. I used the YOLO model to get the bounding box position as talked about before which the filter then uses to correct estimate. I made it so the model assumes the drone is moving in a constant velocity so the box seems less jittery. In addition i set my noise parameter to 10 pixels between fram to frame to help with the YOLO bounding box jitter. Lastly I started with a high uncertanty before the updates started and made it so after 5 misses the track is dropped.

OUTPUT VIDOES (YOUTUBE) https://www.youtube.com/watch?v=iOmZiaXu5XI https://www.youtube.com/watch?v=LDE9W1MX45A

Green rectangle — YOLO detector bounding box Blue line — full 2D trajectory of Kalman-estimated centres Yellow dot — current Kalman centre (frame had a detection) Orange dot — current Kalman centre (predicted through missed detection)

For the failure cases one of them was fast mostion, I kalman estimates te lag and then catches up to the true position. Even with this there were some parts in video 2 were the drone moved too quick and was so far away from the camera it did not pick it up. The missed detections is handled so that if more than 5 missed frames then the track is dropped completly below then its handled by running without an update. Other failure cases included false negatives and when its out of frame. The false negatives were handled by keeping the predictor running and then it will resume updating on the next possible time. If it was detected to be out of frame it continues to run and pickups up when the next detection is made. I had a very difficult time with video 2 because of the clouds in the backround it was very hard to figure out the right confidence level so it could still detect as a smaller image and not lock on to a dark spot in a cloud. Even with this change I was uncussfull in some areas and the drone was not followed properly due to the backround. In addition I tried to make like a height and width max and min for the tracking but because we want it to work on every video not just the two provided i kept it at .2 which still had problems in video 2.