174 lines
4.2 KiB
Markdown
Executable File
174 lines
4.2 KiB
Markdown
Executable File
# ByteTrack 核心原理
|
||
|
||
> 多目标跟踪（MOT）算法详解
|
||
|
||
---
|
||
|
||
## 1. 问题背景
|
||
|
||
### 传统方法的局限
|
||
|
||
大多数 MOT 方法通过**高于阈值的高分检测框**来获取目标 ID：
|
||
|
||
- 遮挡物体的检测分数较低（低分检测框）
|
||
- 低分物体被直接丢弃
|
||
- 导致**轨迹断裂**和**漏检**
|
||
|
||
### ByteTrack 的解决思路
|
||
|
||
> **核心思想**：不仅仅关联高分检测框，而是关联**所有**检测框
|
||
|
||
---
|
||
|
||
## 2. 算法流程
|
||
|
||
### Two-Stage Association
|
||
|
||
```
|
||
输入：视频序列
|
||
↓
|
||
┌─────────────────────────────┐
|
||
│ Step 1: 高分检测框关联 │
|
||
│ (High Score Detection) │
|
||
│ 使用 IoU/特征匹配高分框 │
|
||
└─────────────┬───────────────┘
|
||
↓
|
||
┌─────────────────────────────┐
|
||
│ Step 2: 低分检测框关联 │
|
||
│ (Low Score Detection) │
|
||
│ 利用轨迹相似度恢复目标 │
|
||
│ 过滤背景误检 │
|
||
└─────────────┬───────────────┘
|
||
↓
|
||
输出：完整轨迹
|
||
```
|
||
|
||
### Byte 名称由来
|
||
|
||
- "Byte" = 8 bits = 256 种可能性
|
||
- 每个检测框有 0-255 的检测分数
|
||
- 关联所有分数的检测框，如同处理字节一样
|
||
|
||
---
|
||
|
||
## 3. 核心创新点
|
||
|
||
### 创新一：双重关联
|
||
|
||
| 阶段 | 输入 | 目的 |
|
||
|------|------|------|
|
||
| **第一关联** | 高分检测框 (score > θ₁) | 找到明确的目标 |
|
||
| **第二关联** | 低分检测框 (θ₂ < score ≤ θ₁) | 恢复被遮挡的目标 |
|
||
|
||
### 创新二：YOLOX 检测器
|
||
|
||
ByteTrack 使用 **YOLOX** 作为检测器：
|
||
|
||
- Anchor-free 设计
|
||
- SimOTA 标签分配
|
||
- 优秀的检测性能
|
||
|
||
### 创新三：简单强大的 Tracker
|
||
|
||
- 不依赖 Re-ID 特征
|
||
- 仅使用 **IoU** 关联
|
||
- 高效且易于部署
|
||
|
||
---
|
||
|
||
## 4. 匹配策略
|
||
|
||
### 卡尔曼滤波
|
||
|
||
- 预测下一帧目标位置
|
||
- 处理物体运动平滑
|
||
|
||
### IoU 匹配
|
||
|
||
```python
|
||
# 高分关联
|
||
matched, unmatched_dets = IoUMatching(tracklets, high_score_dets)
|
||
|
||
# 低分关联（仅对未匹配轨迹）
|
||
matched, unmatched_tracks = IoUMatching(unmatched_tracks, low_score_dets)
|
||
```
|
||
|
||
### 阈值设置
|
||
|
||
| 参数 | 说明 | 典型值 |
|
||
|------|------|--------|
|
||
| `high_score_thresh` | 高分阈值 | 0.5 |
|
||
| `match_thresh` | IoU 匹配阈值 | 0.8 |
|
||
| `low_score_thresh` | 低分阈值 | 0.6 |
|
||
|
||
---
|
||
|
||
## 5. 为什么有效？
|
||
|
||
### 遮挡场景示例
|
||
|
||
```
|
||
帧 1: [Person A] → 检测分数 0.95 ✓
|
||
帧 2: [Person A] → 检测分数 0.85 ✓
|
||
帧 3: [被遮挡] → 检测分数 0.35 ✗ (传统方法丢弃)
|
||
|
||
ByteTrack 处理：
|
||
- 帧 1-2: 正常关联，建立轨迹
|
||
- 帧 3: 低分框仍参与第二关联
|
||
- 利用运动预测 + IoU 匹配
|
||
- 成功恢复被遮挡目标
|
||
```
|
||
|
||
### 效果提升
|
||
|
||
| 指标 | 提升 | 原因 |
|
||
|------|------|------|
|
||
| MOTA | +2-3% | 减少漏检 |
|
||
| IDF1 | +1-10% | 轨迹更连续 |
|
||
| IDs | 大幅减少 | 减少轨迹断裂 |
|
||
|
||
---
|
||
|
||
## 6. 与其他 Tracker 对比
|
||
|
||
ByteTrack 可集成到多种现有 tracker：
|
||
|
||
- ✅ ByteTrack + SORT
|
||
- ✅ ByteTrack + DeepSORT
|
||
- ✅ ByteTrack + MOTDT
|
||
|
||
所有组合都获得 **一致的性能提升**。
|
||
|
||
---
|
||
|
||
## 7. 核心代码框架
|
||
|
||
```python
|
||
class BYTETracker:
|
||
def __init__(self, args):
|
||
self.high_score_thresh = args.track_thresh
|
||
self.low_score_thresh = args.low_thresh
|
||
self.match_thresh = args.match_thresh
|
||
|
||
def update(self, dets, frame_info):
|
||
# 分离高低分检测框
|
||
high_dets = dets[dets[:, 4] >= self.high_score_thresh]
|
||
low_dets = dets[dets[:, 4] < self.high_score_thresh]
|
||
|
||
# 第一阶段：高分关联
|
||
online_targets = self.associate(high_dets, tracks)
|
||
|
||
# 第二阶段：低分关联
|
||
online_targets = self.associate_low(online_targets, low_dets)
|
||
|
||
return online_targets
|
||
```
|
||
|
||
---
|
||
|
||
## 8. 参考
|
||
|
||
- **论文**：[ByteTrack: Multi-Object Tracking by Associating Every Detection Box](https://arxiv.org/abs/2110.06864)
|
||
- **代码**：[FoundationVision/ByteTrack](https://github.com/FoundationVision/ByteTrack)
|
||
- **发表**：ECCV 2022
|