Hanze-Qiu/Assignment_3

Assignment 3: Drone Tracking Dataset & Model

This repository contains the dataset, detector configuration, and Kalman Filter tracking logic for tracking drones in video sequences. The tracking is implemented using a deep learning detector paired with a Kalman filter built via filterpy.

Dataset Choice & Detector Configuration

Dataset

The object detection model was trained using a custom drone dataset (structured in YOLO format via data.yaml). The dataset features frames extracted from videos and focuses extensively on accurate bounding box annotations of drones under varying backgrounds and motion profiles.

Detector Configuration

• Model Architecture: The tracking pipeline uses a trained YOLO architecture (loaded from best.pt), with weights optimized for drone detection (experiments included variants like YOLOv8m, YOLO-26s, and RT-DETR-L).
• Confidence Threshold: A detection confidence threshold of conf = 0.45 is used to filter out low-confidence false positives.
• Frame Extraction: Videos were evaluated at frame rates relative to the drone's speed. Slower tracks were extracted and processed at 2 fps, while faster tracks were processed at 4 fps to optimize between computational overhead and temporal consistency.

---

Kalman Filter State Design & Noise Parameters

To ensure smooth trajectories and handle noisy detections, a discrete Kalman Filter was implemented based on a constant velocity motion model.

State & Measurement Vectors

• State Vector ($x$): Defined as $[x, y, v_x, v_y]^T$, representing the $x$ and $y$ coordinates of the bounding box center alongside their respective velocities, $v_x$ and $v_y$.
• Measurement Vector ($z$): Defined as $[x, y]^T$, which receives the centers of the bounding box detections from the YOLO model.

State Transition & Covariance Matrices

• State Transition Matrix ($F$): Assuming a timestep $dt = 1.0$ (representing 1 frame), the state transitions according to $x' = x + v_x \cdot dt$ and $y' = y + v_y \cdot dt$.
• Measurement Noise Covariance ($R$): Modeled with moderate uncertainty on the detector bounding box center.

  R = [[10,  0],
       [ 0, 10]]
  

• Process Noise Covariance ($Q$): Controls how much the model expects the drone's actual motion to deviate from the constant velocity assumption. Using a base parameter q = 5.0:

  Q = [[5.0,   0,    0,    0],
       [  0, 5.0,    0,    0],
       [  0,   0, 10.0,    0],
       [  0,   0,    0, 10.0]]
  

• Initial Estimation Covariance ($P$): The initial uncertainty is scaled relatively high (P *= 100.0) since the starting velocity is initialized at 0 and needs to be quickly learned from incoming detections.

---

Failure Cases & Handling Missed Detections

Object detection is naturally prone to dropping frames due to motion blur, temporary occlusions, or extreme changes in the drone's angle. The tracker is designed to gracefully handle these missed detections.

Missed Detection Handling ("Coasting")

If the YOLO model fails to detect a drone in a given frame, the Kalman filter relies on its predictive power:

1. Prediction Only: The state is advanced purely using kf.predict(). The update step is skipped since there is no measurement ($z$) to apply.
2. Coast Limit: The pipeline maintains a missed_count. The filter will continue to predict ("coast") for up to a max_coast_limit (configured to 15 frames).
3. Visual Indicator: Coasting frames output a predicted bounding box circle indicator in orange, signaling that the tracker is bridging a gap without hard evidence.

Failure Cases

Despite the coast logic, the tracker can fail or permanently lose the drone under specific scenarios:

1. Prolonged Occlusion: If the drone is hidden or undetected for strictly greater than 15 consecutive frames, the tracker resets and waits for a fresh detection to re-initialize an entirely new Kalman track.
2. Erratic Accelerations: The system assumes a constant velocity model. Severe non-linear drone maneuvers (sharp, unpredicted turns) that occur precisely during a missed detection period will cause the "coasting" trajectory to severely deviate from the drone's actual hidden path.
3. False Positive Hijacking: If the detector produces a high-confidence false positive (like a bird) while the tracker is active, the measurement update step will erroneously shift the Kalman state towards the false positive.