People detection and tracking using depth camera

In this paper we present a method for real-time detection and tracking of people in video captured by a depth camera. For each object to be assessed, an ordered sequence of values that represents the distances between its center of mass to the boundary points is calculated. The recognition is based on the analysis of the total distance value between the above sequence and some pre-defined human poses, after apply the Dynamic Time Warping. This similarity approach showed robust results in people detection. 3 DISTINCTION AND FILTERING OF OBJECTS After the segmentation, we perform the labeling of connected components present in the binary edge image. A unique ID number is assigned to each connected region, contained within a closed edge. This allows us to distinguish the various objects. We count the total number of pixels of each object, and measure the width and height of its bounding box. As we're searching for people, in order to speed up the processing, we ignore all the regions that don't correspond to human body morphological characteristics. Therefore, the object is deleted if: The total number of pixels is smaller than a predefined minimum; The total number of pixels is higher than a predefined maximum; The ratio width / height is higher than 0.8. 4 FEATURE EXTRACTION To distinguish and compare the objects present in depth images, the shape of the object is one of the most expressive features. We apply the Chain Code algorithm [Freeman] to read the position of the boundary points of the object in an orderly manner. The starting point for the Chain Code algorithm is defined by the boundary point located at the maximum value of the x-axis frequency histogram of the top half of the object. Then, the values of the distance between the center of mass of the object and its boundary points, are used as the main feature for identify the object (Fig. 2.d). This step was previously done for the poses that we use in the predefined set for the comparison process, and the arrays were stored in files for posterior use. 5 CLASSIFICATION OF OBJECTS Finally, the descriptor generated in the previous step is compared with the predefined human body poses. To check the similarity of the object being assessed with each of the predefined pose, both arrays are processed with Dynamic Time Warping (DTW) algorithm [Myers]. If the total distance between the two arrays after the DTW is below a minimum preset value, the object is classified as a person (Fig. 2). Depending on the pose to be compared, we have defined different threshold values. If we increase the value in these parameters, the number of poses accepted related to each predefined descriptor will also increase. This can be an advantage to classify different human movements with a low number of predefined poses, but also raises the number of false positives. To avoid this, when the boundary line it's not so expressive, which happens in poses that the legs and arms are close together, the value should be low. Figure 2. Scheme of the proposed approach.