A Novel Human Detection Approach Based on Depth Map via Kinect

標籤：des   style   blog   http   os   io   ar   for   2014   

In this paper, a new method of human detection based on depth map from 3D sensor Kinect is proposed. First, the pixel filtering and context filtering are employed to roughly repair defects on the depth map due to information inaccuracy captured by Kinect. Second, a dataset consisting of depth maps with various indoor human poses is constructed as benchmark. Finally, by introducing Kirsch mask and three-value codes to Local Binary Pattern, a novel Local Ternary Direction Pattern (LTDP) feature descriptor is extracted and is used for human detection with SVM as classifier. The performance for the proposed approach is evaluated and compared with other five existing feature descriptors using the same SVM classifier. Experiment results manifest the effectiveness of the proposed approach.

 

人的檢測

These methods can be roughly divided into three different categories; human model based methods [1], template matching based methods [2] and statistical classification methods [3-5].

LBP (local binary pattern) feature, which is a string of bits obtained by binarizing local neighborhood of pixels with respect to the brightness of central pixel, was recently proposed to capture microscopic local image texture and was applied for human detection successfully

HOG-LBP

LTP (local ternary patterns)

CENTRIST (census transform of histograms)

Unfortunately, since the Kinect mainly depends on speckle method [13], the depth map captured by Kinect often contains much noise. 所以，kinect用這種深度映像檢測人不準確，不穩定

 

Overview of LBP-related features

 

By defining the number of spatial transitions (0/1 changes) in LBP pattern with a U value defined as below in (3), the uniformity of LBP patterns, which refers to the patterns having limited transition or discontinuities (U2) in the circular binary presentation, can be determined, where the U value is given as

The uniform LBP only has 59 bins, one each for 58 possible uniform patterns and one for all of the non-uniform ones.

ternary LTP code

 

In summary, the uniform-LBP reduces the dimensions of LBP, while LTP extends LBP to three-valued codes and therefore enhances its anti-noise performance. The CENTRIST introduces a pyramid structure to LBP and makes a multi-scale observation.

特徵提取

Noise Reduction Filters to Depth Map

Compared with TOF data, the depth map captured by Kinect has mounts of null-value areas, which present as ‘white holes’ in depth map

traditional filters, e.i., mean filtering, Gaussian filtering, usually are utilized to remove salt and pepper noises

The pixel filter is designed to compensate the ‘holes‘, while the context filter is employed to further reduce noise in general.

It should be noted that the frames waiting to be retrieved are limited, because the method ignores the scene change between frames.

The LTDP Feature

LTDP (local ternary direction pattern) is derived from the LBP-related feature descriptors by plugging a specific Kirsch mask [15] and three-valued codes into it

At first, the LTDP is calculated by comparing the relative edge response value of a pixel in different directions. The edge response value (S0~S7)  of a particular pixel is calculated with the Kirsch mask at eight different directions. The masks (M0~M7) are shown in Fig.3.

Second, the 3-valued codes for the eight directions with threshold t are defined as follows

In the experiment of this paper, a uniform pattern argument is designed and a coding scheme is used to split each ternary pattern into its positive and negative halves and subsequently treating them as two separate channels of LTDP features for which separate histograms are computed by combining the results only at the end of the computation.

Histogram of LTDP Feature for Depth Map

a depth map should be divided into several non-overlapping rectangular blocks.

A spatial histogram, concatenating the histograms of all blocks can be employed to represent the whole image.

Classification Algorithm

SVM is relatively robust and easy to be implemented. In this study, both linear kernel SVM and nonlinear kernel SVM are used as classification algorithm.

 

檢測類讀的第一篇paper，同樣也指出了臉部偵測的思路吧~~~

A Novel Human Detection Approach Based on Depth Map via Kinect