# 嵌入式目标检测与实时推流系统

**Repository Path**: chizuruwq/objdetSystem

## Basic Information

- **Project Name**: 嵌入式目标检测与实时推流系统
- **Description**: No description available
- **Primary Language**: C/C++
- **License**: 0BSD
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-08-07
- **Last Updated**: 2024-08-28

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 嵌入式目标检测与实时推流系统
## 初始训练

| 模型      | YOLOv3                |
| --------- | --------------------- |
| framework | darknet               |
| dataset   | kaggle:helmet_dataset |
| epoch     | 50000                 |
| mAP       | 84.16%                |

论文地址：[darknet Git]:https://github.com/AlexeyAB/darknet

```shell
# 初始训练
./darknet detector train cfg/my.data cfg/my-yolov3.cfg  ./darknet53.conv.74
```

## 稀疏训练
- 配置
  - GTX3090
  - tensorflow
  - pytorch
  - python3.7

论文地址：[paper]: https://arxiv.org/abs/1708.06519

稀疏训练400轮

```shell
python train.py --cfg /my.cfg --data /my.data --weights weights/last.pt --epoch 400 --batch-size 16 -sr --s 0.001 --prune 1
```
## 剪枝
- 通道剪枝&层剪枝
```shells 
python3 layer_channel_prune.py --cfg cfg/my.cfg --data /my.data --weights weights/best.pt --shortcuts 12 --global_percent 0.74 --layer_keep 0.1
```
## 蒸馏
```shell
python3 train.py --cfg cfg/best.cfg --data /cfg/my.data --weights weights/prune.weights  --epochs 150 --batch-size 40 --t_cfg cfg/my.cfg --t_weights converted.pt -teacher best.weight
```
## 量化
convert中rknn_transform_416\*416.py
# 硬件说明书
> 3rdparty:第三方库
> build:sh start.sh 生成文件存放地
> include:头文件
> RTSPServer:推理+推流服务器
> best.rknn

- 环境要求
  - OpenCV
  - rga
  - mpp
  - ffmpeg
  - live555

```shell
cd build
./rtsp c 0
vlc rtsp://localhost:8554/mystream
```