Academia.eduAcademia.edu

Figure from:Skeletal Tracking using Microsoft Kinect

See full PDFdownloadDownload figure
In our application, the upper body detector is run to detect the torso. The detected regions are then passed on to the face detector to detect faces. We use a frontal face detector and thus allow only a small angle of tilt for the head. The image is scaled down to half the size and the parameters for the detector are set separately for the face and the upper body detector to speed up the process.

Figure 7 In our application, the upper body detector is run to detect the torso. The detected regions are then passed on to the face detector to detect faces. We use a frontal face detector and thus allow only a small angle of tilt for the head. The image is scaled down to half the size and the parameters for the detector are set separately for the face and the upper body detector to speed up the process.

Related Figures (15)

Figure 2: RGB(left) and Depth(right) images from a Kinect
Figure 2: RGB(left) and Depth(right) images from a Kinect
Figure 3: Overview of the algorithm
Figure 3: Overview of the algorithm
rectagular haar features to classify the region of the image as a positive or negative match.
rectagular haar features to classify the region of the image as a positive or negative match.
The detector uses AdaBoost and constructs a rejection cascade of nodes. Each node in turn is a multitree AdaBoosted classifier with a low rejection rate so that a few number of regions are classified out at each stage. A rejection at anny node terminates the whole cascade. An initial learning phase is used to learn the thresholds t;. Boosting is then used to calculate the weights w;. Finally the function below is used to classify whether the region is in an object of interest or not. Thus effectively a region is classified as a positive if it makes it through the whole cascade(Figure (6).
The detector uses AdaBoost and constructs a rejection cascade of nodes. Each node in turn is a multitree AdaBoosted classifier with a low rejection rate so that a few number of regions are classified out at each stage. A rejection at anny node terminates the whole cascade. An initial learning phase is used to learn the thresholds t;. Boosting is then used to calculate the weights w;. Finally the function below is used to classify whether the region is in an object of interest or not. Thus effectively a region is classified as a positive if it makes it through the whole cascade(Figure (6).
Figure 7: Detected face and upper body regions in the RGB image Figure 6: Rejection cascade of the haar cascade classifier [13]
Figure 7: Detected face and upper body regions in the RGB image Figure 6: Rejection cascade of the haar cascade classifier [13]
Figure 8: Anthropometric data from the NASA Anthropometric Data Source Book width set as twice the face width. The lengths of the arms are fixed at 1.12 times the face width of the subject based on the anthropometric data.
Figure 8: Anthropometric data from the NASA Anthropometric Data Source Book width set as twice the face width. The lengths of the arms are fixed at 1.12 times the face width of the subject based on the anthropometric data.
Figure 10: Extended Distance Transform of a foreground segmented depth image 4.5 Skin Segmentation
Figure 10: Extended Distance Transform of a foreground segmented depth image 4.5 Skin Segmentation
Figure 11: Skin segmented image
Figure 11: Skin segmented image
Figure 12: Illustration of the arm fitting process. The angular search space around the left elbow is shown.
Figure 12: Illustration of the arm fitting process. The angular search space around the left elbow is shown.
The entire algorithm was implemented in Python using the OpenCV [4] and OpenKinect [I libraries. Some of the results are shown in Figure [13] with the skeleton overlayed on the image. The algorithm works well on images with less occlusions. With arm occlusions (Figure (14), the distance transform map becomes a bit unreliable and the algorithm sometimes confuses between the arms for the elbow and wrist joints. As our algorithm assumes arms to be present all the time, poses with full hand occlusions do not produce correct results. A similar situation arises for occluded faces and largely occluded upper bodies. But as the detection is done separately for each frame, the algorithm is able to recover from these errors. Another problem with it is that it sometimes tends to get stuck in local extrema. Poses with projected arm lengths approaching zero, i.e. pointing poses are not well recognized (Figure [15). A rough running time analysis o! the algorithm is shown in Figure {1} Table 1: Running time analysis of the algorithm
The entire algorithm was implemented in Python using the OpenCV [4] and OpenKinect [I libraries. Some of the results are shown in Figure [13] with the skeleton overlayed on the image. The algorithm works well on images with less occlusions. With arm occlusions (Figure (14), the distance transform map becomes a bit unreliable and the algorithm sometimes confuses between the arms for the elbow and wrist joints. As our algorithm assumes arms to be present all the time, poses with full hand occlusions do not produce correct results. A similar situation arises for occluded faces and largely occluded upper bodies. But as the detection is done separately for each frame, the algorithm is able to recover from these errors. Another problem with it is that it sometimes tends to get stuck in local extrema. Poses with projected arm lengths approaching zero, i.e. pointing poses are not well recognized (Figure [15). A rough running time analysis o! the algorithm is shown in Figure {1} Table 1: Running time analysis of the algorithm
Figure 13: Positive results with skeleton overlayed on the RGB image
Figure 13: Positive results with skeleton overlayed on the RGB image
Figure 15: Incorrect skeleton with arms mainly oriented towards the camera axis
Figure 15: Incorrect skeleton with arms mainly oriented towards the camera axis
Figure 14: Incorrect skeleton with arms crossing. Confuses between the arms
Figure 14: Incorrect skeleton with arms crossing. Confuses between the arms