sania808/drone-detections

# Assignment 3 - Drone Detection and tracking

## Overview
Detects and tracks drones in video using a pretrained YOLOv8 model and a Kalman filter.

## Task 1

**Model**: `mshamrai/yolov8n-visdrone` :- YOLOv8n fine-tuned on the VisDrone dataset, loaded via HuggingFace.

**Dataset**: VisDrone

**Confidence threshold**: 0.01 for both videos

**Segmentation strategy**: Detections are grouped into continuous segments.
A frame is only kept if it belongs to a segment of at least 3 consecutive detections, with a gap tolerance of 2 frames

## Task 2
**Noise parameters**:
- R = 10 (measurement noise )
- P = 100 (initial covariance )
- Q = 0.1 (process noise)

## Output Videos
- Video 1: [Frames 1 tracked](https://www.youtube.com/watch?v=Fme5aYoDThI)
- Video 2: [Frames 2 tracked](https://www.youtube.com/watch?v=Xl9HNX7Ycfw)

## Failure Cases
**Video 1**: The model required a very low confidence threshold (0.01) to detect the drone reliably

**Video 2**: A physical camera component (tripod/mount) was visible in the bottom right corner of the frame. This can be fixed by cropping the video before processing or applying an area of interest mask

# 作业3——无人机检测与追踪

## 概述
本作业基于预训练YOLOv8模型与卡尔曼滤波器（Kalman filter）实现视频中的无人机检测与追踪任务。

## 任务1

**模型**：`mshamrai/yolov8n-visdrone`——在VisDrone数据集上微调完成的YOLOv8n模型，通过HuggingFace平台加载。

**数据集**：VisDrone数据集

**置信度阈值**：两段视频均采用0.01作为置信度阈值

**分割策略**：将检测结果分组为连续片段。仅当某一帧属于至少包含3次连续检测的片段时才会保留，且允许的帧间隙容差为2帧。

## 任务2
**噪声参数**：
- R = 10（测量噪声）
- P = 100（初始协方差）
- Q = 0.1（过程噪声）

## 输出视频
- 视频1：[帧1追踪结果](https://www.youtube.com/watch?v=Fme5aYoDThI)
- 视频2：[帧2追踪结果](https://www.youtube.com/watch?v=Xl9HNX7Ycfw)

## 失败案例
**视频1**：该模型需采用极低的置信度阈值（0.01）才能可靠检测到无人机
**视频2**：画面右下角可见实体相机组件（三脚架/安装支架）。可通过在处理前裁剪视频或应用感兴趣区域掩码来修复该问题。