cvtechniques/ParkingLotDetection

# Model Description

**Context**

This YOLOv11 model aims to detect spaces in parking lots, whether filled or empty.

**Training Approach**

Fine-tuned from a YOLOv11 foundation model using Ultralytics framework. Combined from two public parking lot image datasets, standardized and augmented.

**Intended Use Cases**

Potential Use Cases:
- Finding busy times and overall trends in parking lot traffic for urban design
- Traffic monitoring apps
- Parking lot owner monitoring

# Training Data

**Dataset sources**

Data taken from currently two of the biggest public parking lot datasets:

- [Subset of CNRPark-EXT](https://universe.roboflow.com/autonomousparking/cnrpark-ext-cnziv-f7ygy) 1208 images (out of 150,000 total images, info on full dataset can be found [here](http://cnrpark.it/))
- [PKLot](https://public.roboflow.com/object-detection/pklot/1) 12545 images

Combined total 13350 (a few removed due to annotation issues)

**Class distribution**
| 0 (empty parking space) | 1 (filled) |
|--------|--------|
| 395,753 | 358,538 |

Limitation: not all instances of classes are annotated in images; some both empty and filled parking spaces missed (see below)
<img alt="Example Image 1" src="example.png" width="400"/>
<img alt="Example Image 2" src="example2.png" width="400"/>

# Training Procedure

Standardization:

- Resized to letterbox 256x256
- Grayscale
- Auto-Adjust Contrast: Adaptive Equalization

Augmentations (done to simulate snow, fog, more camera angles):

- Rotation: +/- 15°
- Blur: Up to 1.5px
- Noise: Up to 2% px

Split was initally 70-20-10, but train tripled due to annotations:
| Train | Valid | Test |
|--------|----------|------ |
| 28086 | 2537 | 1451 |

Training specs:
- Framework: Ultralytics
- Hardware: Used Google Collab, T14
- 15 Epochs, batch size 64, took .7 hours

# Evaluation Results
<!--
1. **Comprehensive Metrics:** Include all output material from your model training in the file uploads to HuggingFace. In your model card report on the key metrics you feel best represent your model and how it does or does not perform well on your use case.
2. **Detailed Per-Class Breakdown:** Create a comprehensive table showing performance for each class. Include test set counts so users can assess statistical reliability.
3. **Visual Examples of Your Classes:** Include example images showing what each of your classes looks like. This helps users understand what your model was trained to detect/classify. Show representative examples with brief descriptions of distinguishing features.
-->
**Key Visualizations:**

<img alt="Confusion Matrix" src="confusion_matrix_normalized.png" width="700"/>
<img alt="F1-Confidence Graph" src="BoxF1_curve.png" width="700"/>
<img alt="Precision-Confidence Graph" src="BoxP_curve.png" width="700"/>

**Performance Analysis:**
<!--
This is where you demonstrate critical thinking. Interpret your results - don't just present numbers
-->

Model performs extremely well in all metrics, however -

confusion matrix shows false positives with background ~50% of the time…

Possibly a combination of causes:
- Some parking spaces weren’t annotated as show in examples
- Over-standardization or annotations? Possible culprit: 256x256 images+grayscale+Adaptive Equalization = lots of potential for class information loss
- Under-training (Only did 15 epochs)

# Limitations and Biases

**Known failure cases:** Model struggles with distinguishing background from classes. This is a major concern to consider in using this model.

**Data biases:** Biased toward high visibility, parking lots with well defined paint as datasets only contain similar data to this. Additionally, parking lot data from Italy & Brazil only. Other countries' parking lots may appear different.

**Contextual limitations:** Aim for aerial view of parking lot with good visibility.

**Inappropriate use cases:** Bad visibility, weather, non-aerial camera angles. Low contrast, ambiguous environments.

**Ethical considerations:** Check with lot owners before setting up cameras to monitor lots.

**Sample size limitations:** Limitation on snowy, foggy conditions.

# 模型描述

**背景**

本YOLOv11（You Only Look Once v11）模型旨在检测停车场内的车位状态，涵盖空闲与已占用两种场景。

**训练方法**

基于Ultralytics框架，以YOLOv11基础模型为蓝本进行微调。训练数据集由两份公开停车场图像数据集合并而来，并经过标准化处理与数据增强。

**预期应用场景**

潜在应用场景包括：
- 分析停车场车流繁忙时段与整体趋势，辅助城市规划设计
- 集成至交通监控类应用
- 供停车场运营方进行运维监控

# 训练数据

**数据集来源**

当前训练数据取自两份规模最大的公开停车场数据集：
- [CNRPark-EXT子集](https://universe.roboflow.com/autonomousparking/cnrpark-ext-cnziv-f7ygy)：共选取1208张图像（完整数据集包含15万张图像，完整数据集详情可参见[此处](http://cnrpark.it/)）
- [PKLot](https://public.roboflow.com/object-detection/pklot/1)：共12545张图像

合并后总图像量为13350张（因标注问题移除了少量样本）。

**类别分布**
| 0（空闲车位） | 1（已占用车位） |
|--------|--------|
| 395,753 | 358,538 |

**局限性**：数据集中并非所有类别实例均被标注，部分空闲与已占用车位存在漏标情况（详见下文）
<img alt="示例图像1" src="example.png" width="400"/>
<img alt="示例图像2" src="example2.png" width="400"/>

# 训练流程

**标准化处理**：
- 调整至信箱式缩放（letterbox）尺寸256×256
- 转为灰度图像
- 自动调整对比度：采用自适应均衡化

**数据增强**（用于模拟降雪、雾天场景与更多拍摄视角）：
- 旋转：±15°
- 模糊：最大1.5像素
- 添加噪声：最大2%像素级噪声

初始数据集划分为70%-20%-10%的训练-验证-测试集，但因标注问题，训练集样本量扩充为原有的三倍：
| 训练集 | 验证集 | 测试集 |
|--------|----------|------ |
| 28086 | 2537 | 1451 |

**训练参数**：
- 框架：Ultralytics
- 硬件：采用Google Colab平台，搭载T14显卡
- 训练轮次：15轮，批次大小（batch size）为64，总训练时长约0.7小时

# 评估结果

**关键可视化结果**：

<img alt="混淆矩阵" src="confusion_matrix_normalized.png" width="700"/>
<img alt="F1值-置信度曲线" src="BoxF1_curve.png" width="700"/>
<img alt="精确率-置信度曲线" src="BoxP_curve.png" width="700"/>

**性能分析**：

模型在各项指标上均表现优异，但混淆矩阵显示，约50%的假阳性样本来自背景干扰……

可能的成因包括以下几点：
- 部分车位未被标注，如示例所示
- 过度标准化或标注问题？潜在诱因：256×256图像尺寸、灰度化处理与自适应均衡化操作可能导致大量类别特征信息丢失
- 训练不足（仅训练15轮）

# 局限性与偏差

**已知失效场景**：模型难以区分背景与目标类别，这是使用该模型时需重点考量的问题。

**数据偏差**：数据集仅包含高可见度、标线清晰的停车场场景，因此模型偏向于这类数据分布。此外，数据集仅涵盖意大利与巴西的停车场数据，其他国家的停车场外观可能存在差异。

**场景局限性**：模型适用于视野良好的航拍停车场场景。

**不适配应用场景**：低可见度环境、恶劣天气、非航拍视角、低对比度场景或模糊环境。

**伦理考量**：在停车场部署监控摄像头前，需征得停车场运营方同意。

**样本量局限性**：降雪、雾天等特殊场景的训练样本不足。