XIAO ESP32S3之物体检测加入视频流

一、前言

????????由于XIAO ESP32S3开发套件没有显示屏配件，因此加入http视频流功能，可通过浏览器请求ESP32S3上的视频流。

二、思路

1、XIAO ESP32S3启动后通过wifi连接到AP；

2、启动http服务器，注册get_mjpeg处理函数；

3、主任务将算法输出的图像压缩为jpg通过xMessageBuffer传递给get_mjpeg处理函数；

4、连接到同一个AP的终端启动浏览器输入XIAO ESP32S3的IP:8081获取视频流。

三、编写代码

1、加入文件

main文件夹下增加http_stream.cpp、http_stream.h文件

修改CMakeLists.txt文件，加入http_stream.cpp，内容如下：

idf_component_register(SRCS 
    app_main.cpp
    fomo_mobilenetv2_model_data.cpp
    http_stream.cpp
)

2、主函数中加入连接AP和传递图像的功能

修改之后的app_main.cpp代码：

#include <inttypes.h>
#include <stdio.h>

#include "img_converters.h"
#include "core/edgelab.h"
#include "fomo_mobilenetv2_model_data.h"

#include "freertos/FreeRTOS.h"
#include "freertos/task.h"
#include "freertos/message_buffer.h"

#include "http_stream.h"

#define DEMO_WIFI_SSID   "huochaigun"
#define DEMO_WIFI_PASS   "12345678"

#define STORAGE_SIZE_BYTES 128*1024

static uint8_t *ucStorageBuffer;

StaticMessageBuffer_t xMessageBufferStruct;

#define kTensorArenaSize (1024 * 1024)

uint16_t color[] = {
  0x0000,
  0x03E0,
  0x001F,
  0x7FE0,
  0xFFFF,
};

extern "C" void app_main(void) {
    using namespace edgelab;

    Device*  device  = Device::get_device();
    device->init();
    printf("device_name:%s\n", device->get_device_name() );
    
    Network* net = device->get_network();

    el_printf(" Network Demo\n");
    uint32_t cnt_for_retry = 0;
    net->init();
    while (net->status() != NETWORK_IDLE) {
        el_sleep(100);
		if(cnt_for_retry++ > 50) {
            net->init();
            cnt_for_retry = 0;
        }
    }
    el_printf(" Network initialized!\n");

    cnt_for_retry = 0;
    net->join(DEMO_WIFI_SSID, DEMO_WIFI_PASS);
    while (net->status() != NETWORK_JOINED) {
		el_sleep(100);
        if(cnt_for_retry++ > 100) {
            net->join(DEMO_WIFI_SSID, DEMO_WIFI_PASS);
            cnt_for_retry = 0;
        }
	}
    el_printf(" WIFI joined!\n");  

//    Display* display = device->get_display();
    Camera*  camera  = device->get_camera();

 //   display->init();
    camera->init(240, 240);

    ucStorageBuffer = (uint8_t *)malloc( STORAGE_SIZE_BYTES );

    if( ucStorageBuffer != NULL )
    {
        xMessageBuffer = xMessageBufferCreateStatic( STORAGE_SIZE_BYTES, ucStorageBuffer, &xMessageBufferStruct );
        if( xMessageBuffer == NULL )
        {
            // There was not enough heap memory space available to create the
            // message buffer.
            printf("Create xMessageBuffer fail\n");
        }
    }
    else
    {
        printf("malloc ucStorageBuffer fail\n");
    }

    start_http_stream();

    auto* engine       = new EngineTFLite();
    auto* tensor_arena = heap_caps_malloc(kTensorArenaSize, MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT);
    engine->init(tensor_arena, kTensorArenaSize);
    engine->load_model(g_fomo_mobilenetv2_model_data, g_fomo_mobilenetv2_model_data_len);
    auto* algorithm = new AlgorithmFOMO(engine);

    while (true) {
        el_img_t img;
        camera->start_stream();
        camera->get_frame(&img);
        algorithm->run(&img);
        uint32_t preprocess_time  = algorithm->get_preprocess_time();
        uint32_t run_time         = algorithm->get_run_time();
        uint32_t postprocess_time = algorithm->get_postprocess_time();
        uint8_t  i                = 0u;
        for (const auto& box : algorithm->get_results()) {
            el_printf("\tbox -> cx_cy_w_h: [%d, %d, %d, %d] t: [%d] s: [%d]\n",
                      box.x,
                      box.y,
                      box.w,
                      box.h,
                      box.target,
                      box.score);

            int16_t y = box.y - box.h / 2;
            int16_t x = box.x - box.w / 2;
            el_draw_rect(&img, x, y, box.w, box.h, color[++i % 5], 4);
        }
        el_printf("preprocess: %d, run: %d, postprocess: %d\n", preprocess_time, run_time, postprocess_time);
//        display->show(&img);
        uint8_t * jpg_buf;
        size_t jpg_buf_len;

        bool jpeg_converted = fmt2jpg( img.data, img.size, img.width, img.height, PIXFORMAT_RGB565, 30, &jpg_buf, &jpg_buf_len);
        if( jpeg_converted == true )
        {
//            printf("jpg_buf_len:%d\n", jpg_buf_len );

            if( xMessageBuffer != NULL )
            {
                xMessageBufferSend( xMessageBuffer, jpg_buf, jpg_buf_len , 0 );
            }
            
            free(jpg_buf);
        }
        camera->stop_stream();
    }

    delete algorithm;
    delete engine;
}

说明：RGB转为jpg的图像质量为30，这样转换后的图形数据小，视频流更流畅。

3、http服务

http_stream.cpp代码：

#include <stdio.h>
#include <stdint.h>
#include <stddef.h>
#include <string.h>

#include "esp_log.h"
#include "esp_timer.h"

#include "freertos/FreeRTOS.h"
#include "freertos/task.h"
#include "freertos/semphr.h"
#include "freertos/event_groups.h"
#include "esp_http_server.h"
#include "freertos/message_buffer.h"


#define PART_BOUNDARY "123456789000000000000987654321"

static const char* _STREAM_CONTENT_TYPE = "multipart/x-mixed-replace;boundary=" PART_BOUNDARY;
static const char* _STREAM_BOUNDARY = "\r\n--" PART_BOUNDARY "\r\n";
static const char* _STREAM_PART = "Content-Type: image/jpeg\r\nContent-Length: %u\r\n\r\n";

static const char *TAG = "http_stream";

MessageBufferHandle_t xMessageBuffer;
/**
  * @brief  jpg_stream_httpd_handler
  * @param  None
  * @retval None
  */
esp_err_t jpg_stream_httpd_handler(httpd_req_t *req)
{

	esp_err_t res = ESP_OK;
	size_t _jpg_buf_size;
  size_t _jpg_buf_len;
	uint8_t * _jpg_buf;
	char * part_buf[64];
	static int64_t last_frame = 0;
	if(!last_frame) 
	{
//		last_frame = esp_timer_get_time();
	}

  _jpg_buf_size = 128*1024;

  _jpg_buf = (uint8_t *)malloc( _jpg_buf_size );

  if( _jpg_buf == NULL )
  {
    return res;
  }

  xMessageBufferReset( xMessageBuffer );
  
	res = httpd_resp_set_type(req, _STREAM_CONTENT_TYPE);
	if(res != ESP_OK)
	{
		return res;
	}

	while(true)
	{

    size_t xReceivedBytes = xMessageBufferReceive( xMessageBuffer, _jpg_buf, _jpg_buf_size, pdMS_TO_TICKS(100) );

    _jpg_buf_len = xReceivedBytes;
    res = ESP_FAIL;

		if( _jpg_buf_len > 0 )
		{
			res = httpd_resp_send_chunk(req, _STREAM_BOUNDARY, strlen(_STREAM_BOUNDARY));
		}
    else
    {
      continue;
    }
		
		if(res == ESP_OK)
		{
			size_t hlen = snprintf((char *)part_buf, 64, _STREAM_PART, _jpg_buf_len);

			res = httpd_resp_send_chunk(req, (const char *)part_buf, hlen);
		}
		

		if(res == ESP_OK)
		{
			res = httpd_resp_send_chunk(req, (const char *)_jpg_buf, _jpg_buf_len);
		}

		if(res != ESP_OK)
		{
			break;
		}

	}

  free( _jpg_buf );
	last_frame = 0;
	return res;
}


httpd_uri_t uri_get_mjpeg = {
    .uri = "/",
    .method = HTTP_GET,
    .handler = jpg_stream_httpd_handler,
    .user_ctx = NULL};

/**
  * @brief  start http_stream.
  * @param  None.
  * @retval None.
  */
httpd_handle_t start_http_stream( void )
{
    printf("start_http_stream\n");

    /* 生成默认的配置参数 */
    httpd_config_t config = HTTPD_DEFAULT_CONFIG();
    config.server_port = 8081;

    /* 置空 esp_http_server 的实例句柄 */
    httpd_handle_t server = NULL;

    /* 启动 httpd server */
    if (httpd_start(&server, &config) == ESP_OK)
    {
        /* 注册 URI 处理程序 */
        // httpd_register_uri_handler(server, &uri_get);
        // httpd_register_uri_handler(server, &uri_get_1m);
        httpd_register_uri_handler(server, &uri_get_mjpeg);
    }
    /* 如果服务器启动失败，返回的句柄是 NULL */
    return server;
}

http_stream.h代码：

#ifndef __HTTP_STREAM_H__
#define __HTTP_STREAM_H__

extern MessageBufferHandle_t xMessageBuffer;

void start_http_stream( void );

#endif

4、修改PSRAM的配置

由于增加的功能需要消耗很多RAM空间，所以需要将PSRAM充分利用起来。

执行idf.py menuconfig，修改SPI RAM config，将malloc()阀值改小一些，这里改为4096字节，意思是使用malloc()分配内存时，大于4096字节则从外部PSRAM获取，配置如下:

?Component config ?--->

????????ESP PSRAM ?--->

????????????????SPI RAM config ?--->

????????????????????????(4096) Maximum malloc() size, in bytes, to always put in internal memory

四?、运行测试

1、编译、烧录、监视

idf.py build
idf.py flash
idf.py monitor

连接AP的日志：

I (813) wifi:mode : sta (dc:54:75:d7:a2:10)
I (813) wifi:enable tsf
 Network initialized!
I WAITING FOR IP...

I (2033) wifi:new:<11,0>, old:<1,0>, ap:<255,255>, sta:<11,0>, prof:1
I (2403) wifi:state: init -> auth (b0)
I (2413) wifi:state: auth -> assoc (0)
I (2423) wifi:state: assoc -> run (10)
W (2423) wifi:[ADDBA]rx delba, code:39, delete tid:5
I (2443) wifi:<ba-add>idx:0 (ifx:0, 90:76:9f:23:a3:58), tid:5, ssn:6, winSize:64
I (2583) wifi:connected with CMCC-2106, aid = 1, channel 11, BW20, bssid = 90:76:9f:23:a3:58
I (2583) wifi:security: WPA2-PSK, phy: bgn, rssi: -63
I (2583) wifi:pm start, type: 1

I (2583) wifi:dp: 1, bi: 102400, li: 3, scale listen interval from 307200 us to 307200 us
I (2583) wifi:set rx beacon pti, rx_bcn_pti: 0, bcn_timeout: 25000, mt_pti: 0, mt_time: 10000
I (2633) wifi:AP's beacon interval = 102400 us, DTIM period = 1
I (3583) esp_netif_handlers: edgelab ip: 192.168.10.104, mask: 255.255.255.0, gw: 192.168.10.1

?可见IP地址为192.168.10.104。

2、浏览器获取视频流

同一个局域网的电脑启动浏览器，地址栏输入:192.168.10.104:8081，然后回车，浏览器会显示视频，图像分辨率为240*240，因为转换为jpg的质量设置的较低，所以不怎么清晰，下图为浏览器显示的图像：

?

有方框表示识别到了物体。

五、总结

此模型主要是对尺寸较小且颜色较深的物体能检测。