md789/Assignment3

Dataset choice and detector configuration. Dataset: Drone Detection Computer Vision Dataset Author: DroneDetectionPITT Dataset Link: https://universe.roboflow.com/dronedetectionpitt-nwyps/drone-detection-yhkcr The dataset contains 34000 drone images across different settings, lighting conditions, and camera proximity. Detector configuration: yolov8. Train 20 epochs. YOLOv8 was chosen because it supports a wide range of tasks, including object detection, that is the main goal of the tasks given.

Kalman filter state design and noise parameters. State (Position and Velocity) = KalmanFilter(dim_x=4, dim_z=2). The 4D state vector (kf.x) represents the horizontal position, vertical position, horizontal velocity, and vertical velocity $[x, y, v_x, v_y]$ of the drone's center point." The state was initialized to be able to track the position and velocity (2D setting). This is important to keep track of the predicted path when the drone speeds up in one direction or changes trajectories. The state transition matrix used was the standard for this task. This initialization was important to keep predicting the next position when the drone was out of frame or when the model could not accurately predict it on the frame. The variable Q, which processes uncertainty and noise, was set to a low value to minimize jumps when tracking the trajectory, resulting in a smoother path. The measurement of uncertainty/noise (R) was set to 5, which is a low value. This tells the filter to have high confidence in the YOLO detections, allowing the tracker to stay tightly locked on the bounding box." The P = 1000 was important to set a high value in the initial uncertainty, so when we start tracking the drone, the initial position is the prediction made by the model and not in the traditional (0,0) coordinates. This resulted in a more accurate initial tracking and more smooth path prediction kf = KalmanFilter(dim_x=4, dim_z=2) kf.x = np.array([0., 0., 0., 0.]) kf.F = np.array([[1, 0, 1, 0], [0, 1, 0, 1], [0, 0, 1, 0], [0, 0, 0, 1]]) kf.H = np.array([[1, 0, 0, 0], [0, 1, 0, 0]]) kf.P *= 1000. kf.R = 5 kf.Q = np.eye(4) * 0.1

Failure cases and how the tracker handles missed detections. Overall, the detector is very consistent when detecting the drone movements, but there are some failure cases that can be detected in the output video. In video 1, at 49 seconds the drone is out of frame, at this time the filter detected the drone comming from the left to center of the frame, this was an error. At 51 seconds the drone is seen appearing again on the top in the center of the frame. Then the filter trackline jumps to track the new apperance of the drone. In this video the filter is dealing well even when the drone is far away in the background. The video 2 is much shorter because only the frames where drone is in frame was kept.
In the source video, the drone is flying more distant from the camera, making it appear smaller, this resulted in some missed predictions, that resulted in less frames and a shorter output video. The filter can track the drone when it's flying high, this creates a contrast between the drone's color and the blue/white from the sky, making it easier to track. However, when the drone is flying near the tree line, away from the camera, the filter is not able to find it. For example, in the source video, around minute 4:18, the drone is flying near the trees, but my filter wasn't able to detect/track it. In this scenario, the trees represent a background noise that was not possible to overcome.

The final videos were composed using OpenCV for the initial tracking overlays and FFMPEG for post-processing. Video 1 = https://youtu.be/Prhxi2QmeYs Video 2 = https://youtu.be/We0wBb3-4s0