文章目錄
Android基於OpenCV通過JNI識別並顯示人臉位置
OpenCV介紹地址:https://docs.opencv.org/2.4/doc/tutorials/introduction/android_binary_package/O4A_SDK.html
Android OpenCV Java Demo地址: https://github.com/kongqw/OpenCVForAndroid,其中人臉比對在Master分支
本文基於JNI實現,源碼地址:Gitee:OpenCVJniFaceDetect
設計思路
取Camera API 中onPreviewFrame 回調的YUV數據送到JNI,
進一步用OpenCV的API識別並畫出人臉區域,再通過ANativeWindow顯示到surface。
代碼設計說明
效果如下
代碼結構如下
其中
app\src\main\cpp\include
來自 opencv-3.4.3-android-sdk\sdk\native\jni\include
app\src\main\assets\lbpcascade_frontalface.xml
來自 opencv-3.4.3-android-sdk\sdk\etc\lbpcascades
app\src\main\jniLibs\
來自 opencv-3.4.3-android-sdk\sdk\native\libs
JNI識別人臉並畫區域代碼如下
// nv21的數據
jbyte *data = env->GetByteArrayElements(data_, NULL);
//mat data->Mat
//1、高 2、寬
Mat src(h + h / 2, w, CV_8UC1, data);
//顏色格式的轉換 nv21->RGBA
//將 nv21的yuv數據轉成了rgba
cvtColor(src, src, COLOR_YUV2RGBA_NV21);
// 可以將Mat的數據寫到存儲卡,正在寫的過程 退出了,導致文件丟失數據
//imwrite("/sdcard/src.jpg",src);
if (cameraId == 1) {
//前置攝像頭,需要逆時針旋轉90度
rotate(src, src, ROTATE_90_COUNTERCLOCKWISE);
//水平翻轉 鏡像
flip(src, src, 1);
} else {
//順時針旋轉90度
rotate(src, src, ROTATE_90_CLOCKWISE);
}
Mat gray;
//灰色
cvtColor(src, gray, COLOR_RGBA2GRAY);
//增強對比度 (直方圖均衡)
equalizeHist(gray, gray);
std::vector<Rect> faces;
//定位人臉 N個
tracker->process(gray);
tracker->getObjects(faces);
for (Rect face : faces) {
//畫矩形 分別指定 bgra
rectangle(src, face, Scalar(255, 0, 0));
}
通過ANativeWindow顯示RGBA數據到surface代碼如下
ANativeWindow_setBuffersGeometry(window, src.cols, src.rows, WINDOW_FORMAT_RGBA_8888);
ANativeWindow_Buffer buffer;
do {
//lock失敗 直接brek出去
if (ANativeWindow_lock(window, &buffer, 0)) {
ANativeWindow_release(window);
window = 0;
break;
}
//src.data : rgba的數據
//把src.data 拷貝到 buffer.bits 裏去
// 一行一行的拷貝
//一行需要多少像素 * 4(RGBA),當stride>width時直接memcpy會顯示異常
//memcpy(buffer.bits, src.data, buffer.stride * buffer.height * 4);
UpdateFrameBuffer(&buffer, src.data);
//提交刷新
ANativeWindow_unlockAndPost(window);
} while (0);
將RGA數據填充到ANativeWindow_Buffer代碼如下
參考自Github:ndk-samples:webp_view.cpp#UpdateFrameBuffer
*
* UpdateFrameBuffer():
* Internal function to perform bits copying onto current frame buffer
* src:
* - if nullptr, blank it
* - otherwise, copy to given buf
* assumption:
* src and bug MUST be in the same geometry format & layout
*/
void UpdateFrameBuffer(ANativeWindow_Buffer *buf, uint8_t *src) {
// src is either null: to blank the screen
// or holding exact pixels with the same fmt [stride is the SAME]
uint8_t *dst = reinterpret_cast<uint8_t *> (buf->bits);
uint32_t bpp;
switch (buf->format) {
case WINDOW_FORMAT_RGB_565:
bpp = 2;
break;
case WINDOW_FORMAT_RGBA_8888:
case WINDOW_FORMAT_RGBX_8888:
bpp = 4;
break;
default:
assert(0);
return;
}
uint32_t stride, width;
stride = buf->stride * bpp;
width = buf->width * bpp;
if (src) {
for (auto height = 0; height < buf->height; ++height) {
memcpy(dst, src, width);
dst += stride, src += width;
}
} else {
for (auto height = 0; height < buf->height; ++height) {
memset(dst, 0, width);
dst += stride;
}
}
}
注意問題說明
- OpenCV 需要依賴 gnustl_static,參考自
opencv-3.4.3-android-sdk\samples\face-detection\jni\Application.mk
, NDK r18b中 移除了gnustl_static,注意選擇NDK17及以下版本 - 目前Camera預覽數據是640X480,在驍龍820手機設備上單幀需要10ms左右,加大尺寸效率會降低
- OpenCV提供的模型對側臉、臉部明暗相差大等情況的識別效果不是特別好