施工场景下违规行为识别数据集

最近，转到了施工场景下的行为识别领域。该应用领域在技术上可拆分为两部分：视频跟踪和行为识别。这一周密集调研了文献，发现着实是一个大坑。其中的视频跟踪最近的各頂会论文出现最多的是单目标跟踪，而我们要解决的确是多目标跟踪，最近出的较好的能实用性的是deepSort；真实的施工场景中摄像头的远近，拍摄的遮挡，工人服装的统一，重叠，违规动作幅度的大小等都是巨大的挑战；行为识别方面最近出的论文较多，能实用性的目前敲定ECO模型；在跟踪过程中某一个工人的时空管道数据的抽取也是一个难题等等。无论如何，这块硬骨头得啃下来。

行为识别模型的训练一般是用摄像头采集某一个动作大概2-3秒的视频，然后将这些小视频打上行为标签用于训练；在测试中得结合视频跟踪出某一个目标，再抽取该目标的时空管道数据，主要数据的抽取中背景也很重要。

接下来，采集数据。带着4个工大本科生开启了为期10天的数据采集体验。为了提神效率，笔记本搭载了3个仅有的摄像头不同角度的抓取数据。充分考虑光照变化、场景变化、工人高低、行为夸张、拍摄角度等，进行不同组合来抓取行为视频数据。感谢4个本科生，这10天共收集了640个小视频以及测试用的demo。
抓取数据借用OpenCV，的代码如下：

#include "opencv2/opencv.hpp"
#include <iostream>
#include <iomanip>  

#include <windows.h>
#define KEY_DOWN(VK_NONAME) ((GetAsyncKeyState(VK_NONAME) & 0x8000) ? 1:0) //必要

using namespace std;
using namespace cv;

int main() {

	// Create a VideoCapture object and use camera to capture the video
	VideoCapture cap1(0);
	VideoCapture cap2(1);

	VideoCapture cap3(2);
	// Check if camera opened successfully
	if (!cap1.isOpened() || !cap2.isOpened() || !cap3.isOpened())
	{
		cout << "Error opening video stream" << endl;
		return -1;
	}

	// Default resolution of the frame is obtained.The default resolution is system dependent. 
	//int frame_width = cap1.get(CV_CAP_PROP_FRAME_WIDTH);
	//int frame_height = cap1.get(CV_CAP_PROP_FRAME_HEIGHT);
	 
	string filename;
	string cam_num1 = "cam1";
	string cam_num2 = "cam2";
	string cam_num3 = "cam3";
	
	while (1){

		//filename << "smoking32";   //smoking27   wear_hat27
		//cam_num1 << "cam7";
		//cam_num2 << "cam8";
		//cam_num3 << "cam9";
		if (!cap1.isOpened() || !cap2.isOpened() || !cap3.isOpened())
		{
			cout << "Error opening video stream" << endl;
			break;
		}

		//cout << "please input action and cam_num." << endl << "For example:" << endl << "smoking32" << endl << "cam7" << endl << "cam8" << endl << "cam9" << endl;
		//cin >> filename >> cam_num1 >> cam_num2 >> cam_num3;

		cout << "please input action and cam_num." << endl << "For example:" << endl << "smoking32" << endl << endl;
		cout << "Now, please input (input 'q' will exit):" << endl;
		cin >> filename;

		if (filename == "q")
			break;

		// Define the codec and create VideoWriter object.The output is stored in 'outcpp.avi' file. 
		VideoWriter video1("C:\\个人\\dataset\\ASD\\" + filename + "\\" + cam_num1 + "\\" + filename + cam_num1 + ".avi", CV_FOURCC('D', 'I', 'V', 'X'), 30, Size(640, 480));
		VideoWriter video2("C:\\个人\\dataset\\ASD\\" + filename + "\\" + cam_num2 + "\\" + filename + cam_num2 + ".avi", CV_FOURCC('D', 'I', 'V', 'X'), 30, Size(640, 480));
		VideoWriter video3("C:\\个人\\dataset\\ASD\\" + filename + "\\" + cam_num3 + "\\" + filename + cam_num3 + ".avi", CV_FOURCC('D', 'I', 'V', 'X'), 30, Size(640, 480));


		long currentFrame = 1;

		while (1)
		{
			Mat frame1;
			Mat frame2;
			Mat frame3;
			// Capture frame-by-frame 
			cap1 >> frame1;
			cap2 >> frame2;
			cap3 >> frame3;

			// If the frame is empty, break immediately
			if (frame1.empty() || frame2.empty() || frame3.empty())
				break;


			stringstream str1;
			str1 << "img_" << setw(4) << setfill('0') << currentFrame << ".jpg";
			cout << "正在处理第" << currentFrame << "帧" << endl;
			printf("\n");
			// 设置每1帧获取一次帧
			if (currentFrame % 1 == 0) {
				// 将帧转成图片输出
				imwrite("C:\\个人\\dataset\\ASD\\" + filename + "\\" + cam_num1 + "\\" + str1.str(), frame1);
				imwrite("C:\\个人\\dataset\\ASD\\" + filename  + "\\" + cam_num2 + "\\" + str1.str(), frame2);
				imwrite("C:\\个人\\dataset\\ASD\\" + filename + "\\" + cam_num3 + "\\" + str1.str(), frame3);
			}


			currentFrame++;
			// Write the frame into the file 'outcpp.avi'
			video1.write(frame1);
			video2.write(frame2);
			video3.write(frame3);

			// Display the resulting frame    
			imshow("cam1", frame1);
			imshow("cam2", frame2);
			imshow("cam3", frame3);
			waitKey(5);

			if (KEY_DOWN(32))
				break;

			if (KEY_DOWN(27))
			{
				video1.release();
				video2.release();
				video3.release();
				cap1.release();
				cap2.release();
				cap3.release();
				// Closes all the windows
				destroyAllWindows();
				return 0;
			}
		}


		// When everything done, release the video capture and write object
		
		video1.release();
		video2.release();
		video3.release();	
	}
	
	cap1.release();
	cap2.release();
	cap3.release();
	// Closes all the windows
	destroyAllWindows();
	return 0;
}