Research on intelligent substation monitoring by image recognition method

3.1 Background difference method

The key point of the background subtraction method is to construct an appropriate background model. In this study, the background model is constructed by Gaussian mixture distribution, and the steps are as follows.

1. Firstly, the first few frames of the video are taken as the sampling samples to construct the initial Gaussian mixture background model [10]. In the mixture Gaussian background model, the probability density function of pixels conforming to the Gaussian mixture distribution is as follows:

where p(xt) stands for the image pixel probability density at time t, xt stands for the image pixel sample at time t, k represents the number of superimposed Gaussian models, ωi,t represents the weight of the i-th Gaussian distribution of image pixel at time t, η(xt,μi,t,τi,t) represents the i-th Gaussian distribution of image pixel at time t, μi,t represents the mean value of the corresponding Gaussian distribution, τi,t is the corresponding covariance matrix, δi,t is the corresponding variance, and I is the three-dimensional identity matrix of RGB pixels [11].

1. Differential comparison was performed on the pixel value of the latest frame image and k Gaussian distribution models that are established in the previous step, according to Equation (2). If it is consistent, it is the background; otherwise, it is the moving target.

1. The weight of the model is updated, and the models are matched. The update formula is as follows:

where α is the learning rate [12], Mi,t is one when the pixel point in the difference conforms to the i-th model and zero when the pixel point does not conform to the i-th model, η(xt|μi,σi) stands for the i-th Gaussian distribution of the image pixel at time t, μi and σi are the mean value and standard deviation of the corresponding Gaussian distribution respectively, and ρ is the updated parameter.

The above process is repeated to constantly compare the new frame image with the background model to find out the pixels that do not match the background model and taken them as the pixels of the moving target, and the background model is constantly updated to ensure the difference accuracy.

3.2 Frame difference method

3.2.1 Three-frame difference method

The frame difference rule obtains the brightness difference through the difference between successive frames and determines the pixel of the moving target according to the brightness difference [13]. The background in the image is mostly static compared with the moving object, and it often changes little in a short period in two successive frames. However, when the moving object moves rapidly in the image, its pixel will change obviously, thus the moving region can be found through comparison. However, when only comparing two successive images, it is not only easy to make the extracted object appear “ghost shadow”, but also difficult to detect the overlapping part of targets. Therefore, this study applied the three-frame difference method to make up for the above shortcomings.

The processing of video by the three-frame difference method is shown in Figure 2. The gray difference is performed on the current frame and the last and next frames, respectively, and then the two difference images are binarized and superimposed with logic “AND” to obtain the moving object in the video.

Figure 2:

The recognition process of moving target by the three-frame difference method in video images.

3.2.2 Improved three-frame difference method

Compared with the traditional frame difference method, the three-frame difference method does not only compare the difference between two adjacent frames but also makes the current frame superimposed with the last and next frames after differential binarization. Although the three-frame difference method can eliminate the “ghost shadow”, when facing the low-speed target, there is no change in the short time between the two frames, and the overlapping part will also produce “cavities”. To eliminate the above phenomenon, this study improved the three-frame difference method. The current frame is no longer differential with the last and next frames, but with the background model and next frame. There are many methods to establish a background model. Due to the large data of substation video monitoring and the timely identification of suspicious personnel, a relatively simple mean background modeling [14] was adopted in this study. The flow chart is shown in Figure 3.

1. Firstly, the video is read frame by frame and grayed.

2. k=1 is set. The kth frame image is selected for moving target recognition. Firstly, the average background of the k1-frame image is modeled.

(4)avgk=f1+f2+f3+⋯+fkk,

where fk is the gray image of the kth frame and avgk is the average background model of the kth frame.

1. The kth frame gray image treated by differential processing respectively with the background model and the k+1-th frame gray image, respectively, and the formulas are:

(5){Mk,1=|fk−avgk|Mk,2=|fk+1−avgk|,

where Mk,1 is the differential between the kth frame and the background model, and Mk,2 is the differential between the k + 1 frame and the background model.

1. The differential image is binarized; then the two binary differential images are overlapped in the form of logical “AND”, and the overlapped images are denoised by the mean filter to obtain the moving target recognition result of the kth frame.

2. Whether k is smaller than m − 1 or not is determined, where m is the total number of video frames. If it is, then k = k + 1, and it returns to Step (2); if not, the processed video image is output.

Figure 3:

The flow of video processing after improvement.

4.1 Experimental environment

In this study, the improved three-frame difference method was simulated by MATLAB software [15]. The experiment was conducted in a laboratory server whose configurations were Windows7 system, I7 processor, and 16 G memory.

4.2 Experimental data

The video used in the simulation experiment was from three surveillance cameras of a substation in Shanghai. These cameras were respectively set at the entrance and exit of the substation and the important equipment placement area. Moreover, the video extracted from the three cameras has been checked to ensure that there were enough moving targets in the captured video time, mainly the workers in and out. The basic parameters of videos extracted by the cameras are shown in Table 1.

4.3 Experimental project

The moving target recognition was carried out on the above three videos using three recognition methods. The learning rate was set as 0.02 when the Gaussian mixture background method updated the background model; the traditional and improved three-frame difference methods needed to binarize the differential image, and the threshold of binarization was the average value of the current differential image.

In addition to the above simulation experiments, the three moving target detection algorithms were tested in reality. The test was carried out in the substation from which the video used in the simulation experiment was obtained. The three algorithms were used in the unattended monitoring of the substation. The test lasted for one month. To test the effect of the three algorithms, although it is an unsupervised substation, personnel were still dispatched to assist the monitoring. The staffs who were responsible for recording worked in three shifts, 8 h each shift. The staffs recorded the actual number of moving targets in and out of the monitoring video and times of illegal invasion.

4.4 Experimental results

The total number of monitoring video frames used in the simulation experiment was large. Limited by the space, only one frame of the original image and the moving object detection results processed by the three moving target detection algorithms were displayed. As shown in Figure 4, (1) is a frame of image in the original video, in which two substation staffs are inspecting the substation equipment, and the background mainly includes substation equipment, roads, sky, and single-layer building; (2) is the result of the Gaussian mixture background model method, most of the background is removed, and there are basic outlines of the staffs, but the residual contour of the single-layer building interferes with the staffs, and the staffs only have the basic contour, with cavities in the middle; (3) is the processing result of the traditional three-frame difference method, which only retains the outline of the staffs in the original image, with large cavities in the middle; (4) is the result of the improved method, which highlights the staffs and has no cavities.

Figure 4:

Moving target detection results of three algorithms after processing videos.

In this study, the performance of the three algorithms in detecting moving a target in the simulation video was measured by the precision and recall ratios. As shown in Figure 5, the average precision and recall ratios of the Gaussian mixture background model method were 66.3 and 51.0% respectively, the average precision and recall ratios of the traditional three-frame difference method were 87.6 and 78.5% respectively, and the average precision and recall ratios of the improved three-frame difference method were 95.3 and 88.9% respectively.

Figure 5:

The detection performance of three algorithms for simulation video.

In addition to the above simulation experiments, this study also applied the three algorithms to the video monitoring of substation to test the effect of their practical application, and the application effect is shown in Table 2. The actual times of illegal invasion in the substation in one month was obtained by manual monitoring and recording, with a total of 76 times; 70 times were wild animals that intruded into the substation facilities by mistake, four times out of the remaining six times were that ordinary people enter the monitoring range outside the substation due to being lost, and only two times were illegal elements deliberately entering the substation to carry out illegal acts. For the monitoring of illegal intrusion in the substation, the accurate times and the number of false positives under the Gaussian mixture background model method was 53 and 25 respectively, the accurate times and the number of false positives under the traditional three-frame difference method were 69 and 9, and the accurate times and the number of false positives of the improved three-frame difference method were 75 and 2 respectively.

The above experimental results reflected the recognition effect of the three algorithms for the moving targets in the actual intelligent substation. The monitoring record results demonstrated that the invasion times of illegal elements was small among the detected illegal intrusion events, and the entering of wild animals by mistake was the most common. As the intelligent substation is often in remote areas, it is easier to be targeted by criminals in theory, but the terrain in the remote area is complex, which makes the invasion of wild animals common. To effectively distinguish between wild animals and regular staff, in the actual substation test in this study, the body size of the regular staff in the substation was recorded, and the upper and lower limits of the size were established. The size of wild small animals is often different from that of the human body. Therefore, after processing the monitoring image, the three algorithms estimated the body shape of the extracted moving objects. Once the upper and lower limits were exceeded, records and alarms would be made. The accuracy of these algorithms was directly affected by the accuracy of the three moving target recognition algorithms, whether it was the shape estimation or the comparison of the upper and lower limits. The simulation results demonstrated that the contour of the moving target processed by the improved three-frame difference method was the clearest, with the least background interference; therefore, the interference to the body size estimation of the moving target was the least, and the number of false alarms was also the least.

Among the three moving target recognition algorithms, the traditional three-frame difference method only needed to compare the current frame with the upper and lower frames, thus the calculation amount was small. The background model needed to be established when the Gaussian mixture background method was used for difference, and the background model needed to be updated continuously in this process, leading to a large calculation amount. The calculation amount was small when the improved three-frame difference method performed difference on the next frame, but when performing difference on the background model calculated by the Gaussian mixture background method, the calculation of the background model accounted for a very large proportion; therefore, the calculation amount of the improved three-frame difference method was larger than that of the traditional three-frame difference method.