1*4882a593Smuzhiyun# Yolo-v5 demo 2*4882a593Smuzhiyun 3*4882a593Smuzhiyun## 导出rknn模型 4*4882a593Smuzhiyun 5*4882a593Smuzhiyun请参考 https://github.com/airockchip/rknn_model_zoo/tree/main/models/vision/object_detection/yolov5-pytorch 6*4882a593Smuzhiyun 7*4882a593Smuzhiyun 8*4882a593Smuzhiyun 9*4882a593Smuzhiyun## 注意事项 10*4882a593Smuzhiyun 11*4882a593Smuzhiyun1. 使用rknn-toolkit2版本大于等于1.1.2。 12*4882a593Smuzhiyun2. 切换成自己训练的模型时，请注意对齐anchor等后处理参数，否则会导致后处理解析出错。 13*4882a593Smuzhiyun3. 官网和rk预训练模型都是检测80类的目标，如果自己训练的模型,需要更改include/postprocess.h中的OBJ_CLASS_NUM以及NMS_THRESH,BOX_THRESH后处理参数。 14*4882a593Smuzhiyun5. demo需要librga.so的支持,编译使用请参考https://github.com/rockchip-linux/linux-rga 15*4882a593Smuzhiyun5. 由于硬件限制，该demo的模型默认把 yolov5 模型的后处理部分，移至cpu实现。本demo附带的模型均使用relu为激活函数，相比silu激活函数精度略微下降，性能大幅上升。 16*4882a593Smuzhiyun 17*4882a593Smuzhiyun 18*4882a593Smuzhiyun 19*4882a593Smuzhiyun## Android Demo 20*4882a593Smuzhiyun 21*4882a593Smuzhiyun### 编译 22*4882a593Smuzhiyun 23*4882a593Smuzhiyun根据指定平台修改 `build-android_<TARGET_PLATFORM>.sh`中的Android NDK的路径 `ANDROID_NDK_PATH`，<TARGET_PLATFORM>可以是RK356X或RK3588 例如修改成： 24*4882a593Smuzhiyun 25*4882a593Smuzhiyun```sh 26*4882a593SmuzhiyunANDROID_NDK_PATH=~/opt/tool_chain/android-ndk-r17 27*4882a593Smuzhiyun``` 28*4882a593Smuzhiyun 29*4882a593Smuzhiyun然后执行： 30*4882a593Smuzhiyun 31*4882a593Smuzhiyun```sh 32*4882a593Smuzhiyun./build-android_<TARGET_PLATFORM>.sh 33*4882a593Smuzhiyun``` 34*4882a593Smuzhiyun 35*4882a593Smuzhiyun### 推送执行文件到板子 36*4882a593Smuzhiyun 37*4882a593Smuzhiyun连接板子的usb口到PC,将整个demo目录到 `/data`: 38*4882a593Smuzhiyun 39*4882a593Smuzhiyun```sh 40*4882a593Smuzhiyunadb root 41*4882a593Smuzhiyunadb remount 42*4882a593Smuzhiyunadb push install/rknn_yolov5_demo /data/ 43*4882a593Smuzhiyun``` 44*4882a593Smuzhiyun 45*4882a593Smuzhiyun### 运行 46*4882a593Smuzhiyun 47*4882a593Smuzhiyun```sh 48*4882a593Smuzhiyunadb shell 49*4882a593Smuzhiyuncd /data/rknn_yolov5_demo/ 50*4882a593Smuzhiyun 51*4882a593Smuzhiyunexport LD_LIBRARY_PATH=./lib 52*4882a593Smuzhiyun./rknn_yolov5_demo model/<TARGET_PLATFORM>/yolov5s-640-640.rknn model/bus.jpg 53*4882a593Smuzhiyun``` 54*4882a593Smuzhiyun 55*4882a593Smuzhiyun## Aarch64 Linux Demo 56*4882a593Smuzhiyun 57*4882a593Smuzhiyun### 编译 58*4882a593Smuzhiyun 59*4882a593Smuzhiyun根据指定平台修改 `build-linux_<TARGET_PLATFORM>.sh`中的交叉编译器所在目录的路径 `TOOL_CHAIN`，例如修改成 60*4882a593Smuzhiyun 61*4882a593Smuzhiyun```sh 62*4882a593Smuzhiyunexport TOOL_CHAIN=~/opt/tool_chain/gcc-9.3.0-x86_64_aarch64-linux-gnu/host 63*4882a593Smuzhiyun``` 64*4882a593Smuzhiyun 65*4882a593Smuzhiyun然后执行： 66*4882a593Smuzhiyun 67*4882a593Smuzhiyun```sh 68*4882a593Smuzhiyun./build-linux_<TARGET_PLATFORM>.sh 69*4882a593Smuzhiyun``` 70*4882a593Smuzhiyun 71*4882a593Smuzhiyun### 推送执行文件到板子 72*4882a593Smuzhiyun 73*4882a593Smuzhiyun 74*4882a593Smuzhiyun将 install/rknn_yolov5_demo_Linux 拷贝到板子的/userdata/目录. 75*4882a593Smuzhiyun 76*4882a593Smuzhiyun- 如果使用rockchip的EVB板子，可以使用adb将文件推到板子上： 77*4882a593Smuzhiyun 78*4882a593Smuzhiyun``` 79*4882a593Smuzhiyunadb push install/rknn_yolov5_demo_Linux /userdata/ 80*4882a593Smuzhiyun``` 81*4882a593Smuzhiyun 82*4882a593Smuzhiyun- 如果使用其他板子，可以使用scp等方式将install/rknn_yolov5_demo_Linux拷贝到板子的/userdata/目录 83*4882a593Smuzhiyun 84*4882a593Smuzhiyun### 运行 85*4882a593Smuzhiyun 86*4882a593Smuzhiyun```sh 87*4882a593Smuzhiyunadb shell 88*4882a593Smuzhiyuncd /userdata/rknn_yolov5_demo_Linux/ 89*4882a593Smuzhiyun 90*4882a593Smuzhiyunexport LD_LIBRARY_PATH=./lib 91*4882a593Smuzhiyun./rknn_yolov5_demo model/<TARGET_PLATFORM>/yolov5s-640-640.rknn model/bus.jpg 92*4882a593Smuzhiyun``` 93*4882a593Smuzhiyun 94*4882a593SmuzhiyunNote: Try searching the location of librga.so and add it to LD_LIBRARY_PATH if the librga.so is not found on the lib folder. 95*4882a593SmuzhiyunUsing the following commands to add to LD_LIBRARY_PATH. 96*4882a593Smuzhiyun 97*4882a593Smuzhiyun```sh 98*4882a593Smuzhiyunexport LD_LIBRARY_PATH=./lib:<LOCATION_LIBRGA.SO> 99*4882a593Smuzhiyun``` 100*4882a593Smuzhiyun 101*4882a593Smuzhiyun## 视频流Demo运行命令参考如下： 102*4882a593Smuzhiyun- h264视频 103*4882a593Smuzhiyun``` 104*4882a593Smuzhiyun./rknn_yolov5_video_demo model/<TARGET_PLATFORM>/yolov5s-640-640.rknn xxx.h264 264 105*4882a593Smuzhiyun``` 106*4882a593Smuzhiyun注意需要使用h264码流视频，可以使用如下命令转换得到： 107*4882a593Smuzhiyun``` 108*4882a593Smuzhiyunffmpeg -i xxx.mp4 -vcodec h264 out.h264 109*4882a593Smuzhiyun``` 110*4882a593Smuzhiyun 111*4882a593Smuzhiyun- h265视频 112*4882a593Smuzhiyun``` 113*4882a593Smuzhiyun./rknn_yolov5_video_demo model/<TARGET_PLATFORM>/yolov5s-640-640.rknn xxx.hevc 265 114*4882a593Smuzhiyun``` 115*4882a593Smuzhiyun注意需要使用h265码流视频，可以使用如下命令转换得到： 116*4882a593Smuzhiyun``` 117*4882a593Smuzhiyunffmpeg -i xxx.mp4 -vcodec hevc out.hevc 118*4882a593Smuzhiyun``` 119*4882a593Smuzhiyun- rtsp视频流 120*4882a593Smuzhiyun``` 121*4882a593Smuzhiyun./rknn_yolov5_video_demo model/<TARGET_PLATFORM>/yolov5s-640-640.rknn <RTSP_URL> 265 122*4882a593Smuzhiyun``` 123*4882a593Smuzhiyun 124*4882a593Smuzhiyun### 注意 125*4882a593Smuzhiyun 126*4882a593Smuzhiyun- 需要根据系统的rga驱动选择正确的librga库，具体依赖请参考： https://github.com/airockchip/librga 127*4882a593Smuzhiyun- **rk3562 目前仅支持h264视频流** 128*4882a593Smuzhiyun- **rtsp 视频流Demo仅在Linux系统上支持，Android上目前还不支持**
