Group of Fragmental Regions

In some cases, the detected foreground regions are fragmental. This is because the color similarity between the dress of person and background pixel is too low to differentiate them. As shown in Fig. 2.2.1, we will obtain fragmental foreground regions in some situations.

Fig. 2.2.1 Examples of fragmental regions cases.

In this section, we propose a method to solve this problem. We use the following

flow diagram to depict how we fill the fragmental foreground regions.

Fig. 2.2.2 The flow diagram of grouping fragmental regions.

In 2.2.1, we use edge detection to find the contour of a person. By finding the contour of the person, we can regain the missing regions. However there are too many edges to be detected when detecting edges, so we use background edges to filter those unneeded edges in subsection 2.2.2. Then we can locate the regions of missing persons. Finally, in subsection 2.2.3, we fill the fragmental regions of the missing parts to obtain the person region.

2.2.1 Edge detection

We detect edges because of the low color similarity between dress of person and background pixel. A portion of human bodies are missed by a pixel-wise differencing method (the method of modified frame differencing). However, the contour of a human can be detected by a template-based edge detector.

Fig. 2.2.1.2 The results of using pixel-wise differencing method.

The above figures show us the results of a pixel-wise differencing method in some cases of low color similarity. We will obtain the contour of the person to fill the missing parts of the body. We will use edge information [20] to obtain the contour of the person.

Because the Prewitt edge detector is more sensitive to edges than other edge detectors, we use it to detect edges. Additionally, we choose the vertical edges to locate the position of the person since most man-made buildings contain many vertical edges and human shape contains more vertical edges than horizontal ones. Fig.

2.2.1.3 shows the results of vertical and horizontal edge detectors.

Fig. 2.2.1.3 The results of vertical and horizontal edge detections.

Although the detected results of the foreground regions are fragmental, it can be used to know the positions of the fragmental regions to be filled.

2.2.2 Background edge removal

There are many edges detected as shown in Fig. 2.2.2.1. The edges belonging to the person are our targets. The remaining edges need to be cleared, which include all non-human edges. In this section, we present our method to clear those redundant edges.

We use the background edges belong to several frames to clear the background edge of the current frame. To avoid the slow movement of the person, the interval between those frames and the time of the current frames must be long enough. In other words, if the person moves slowly and the interval is not far enough, we may clear the target edges erroneously. As shown in Fig. 2.2.2.2, we clear the background edges of the current frame by using picked frames.

Fig. 2.2.2.1 Images showing our background edges removal method.

These marked frames are used to clear the background of the current frame

After clearing the background edges, the contour of person can be obtained and we will fill the lost foreground regions. The results after applying the method of this section are shown on the following figure:

Fig. 2.2.2.3 The results of our background edges removal method.

2.2.3 Lost Foreground Filling

After the background edges were cleared, we use the information of detected foreground to locate the range of regions we want to fill. As shown in Fig. 2.2.3.1, we check a range near the person. In this range, we detect edge points between the left and right boundary of the range in each row. We then find a left and a right edge point in each row and define the foreground run as the closed interval between the left and right edge points. Finally, we fill each foreground run to obtain the region of the person.

Fig. 2.2.3.1 A diagram showing our background edges removal method.

Fig. 2.2.3.2 depicts the result after using our method.

Fig. 2.2.3.2 The results of our lost foreground filling method.

2.3 Multiple Persons Segmentation

Considering the case of connect persons, the aspect ratio of the foreground region does not meet the ratio of a person. If the aspect ratio of a foreground region exceeds a threshold, then this region may contain multiple persons. As shown in Fig 2.3.1, we try to find a vertical line to split the region of multiple persons.

Fig. 2.3.1 An example of connected persons case.

In this section we present a method to split the detected region of connected persons.

The following diagram shows how we segment a region of connected persons.

Fig. 2.3.2 The flow diagram of the multiple persons segmentation algorithm.

Here we introduce our method to split the region of the connected persons briefly.

Firstly, we define a vertical scanning line and find the distance between the top and bottom edge points in the edge map for each vertical scanning line. Second, we define the separating line as the vertical scanning line with the shortest distance. As shown in Fig. 2.3.3, the vertical blue line is the separating line of this edge map.

Fig. 2.3.3 A diagram to show the separating line.

CHAPTER 3 HUMAN BODY DECOMPOSITION