Top-view people detection based on multiple subarea pose models for smart home system

Abstract

In this paper, an effective top-view people detection algorithm based on multiple subarea models is proposed for smart home system. Conventional single model based detector is difficult to achieve high performance in top-view people detection since there are too many possible individual poses in the top-view based image scene and it is impossible to cover all the poses with single model. Therefore, this paper develops a model of 9 typical poses to mitigate the low detection performance problem of conventional method. Moreover, by restricting the local scope of every pose model, the proposed approach yields an improved detection rate while reducing false alarm compared to the conventional single model based detector.

Bipolar Log‐Intensity‐Variance Histogram Method for Local Image Patch Intensity Change Measurements

“…For outdoor scene motion change DB4, 250 80-by-80 local patch pairs were extracted from an outdoor people-walking scene image sequence [30], as Figure 7(d) shows. For indoor scene motion change DB5, 290 80-by-80 local patch pairs were extracted from a top-view indoor people-walking scene image sequence [20] as Figure 7(e) shows. Therefore, our test local image patch pair DB consisted of five subdatabases: indoor illumination change DB1, outdoor illumination change DB2, indoor motion change DB3 and DB5, and outdoor motion change DB4.…”

“…For a fixed-position camera, the intensity changes within two local adjacent image patches are always caused by illumination changes or object movement. To reduce false alarms caused by illumination changes for accurate motion- [1][2][3][4][5][6] or change-detection [7][8][9][10][11][12][13][14][15], as conventional methods, many robust feature descriptors have been proposed [16][17][18][19][20]. For example, a local binary pattern (LBP) [21], which checks the relative difference between spatially neighboring pixels and not the absolute values of each pixel, illumination changes can be effectively overcome.…”

“…Song et al [26] proposed a box-count dimension measure based on CPPS to reduce lip-motion detection false alarms caused by illumination changes. Similarly, Wang et al [20] applied CPPS and the box-count dimension measure method to enhance top-view moving people detection system performance.…”

“…Although CPPS has been used as a feature space to construct metric functions of intensity changes, existing work on CPPS and box-count dimension measurements [20,25,26] has some notable shortcomings. First, the box-count dimension measure lacks a facility for deep analysis of the CPPS property under different types of intensity changes.…”

“…Illumination change can also cause nonlinear pixel changes, as shown in Figure 1(a). As a result, CPPS and box-count measurement methods [20,25,26] fail in many cases, such as slight motion and strong illumination changes.…”