1. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. High level representations of the two data flows. (a) The state-of-the-art data flow of the OpenPTrack system on the RGB-D smart camera. The 3-D point cloud is computed outside the libfreenect2 library by the ROS nodelet cloud_generation_node. (b) The proposed data flow of the same system. The 3-D point cloud is now directly computed within the new version of libfreenect2 and streamed by the new version of kinect2_bridge, thus the external nodelet is no longer required. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007

2. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. The processing flow performed by our library for each single frame. At start-up, the space for the needed data is allocated once in the GPU memory, then, for each frame, the data coming from the CPU memory are processed to achieve the final data. The data transferred from CPU to GPU consist of the raw IR and depth data, while the data transferred from GPU to CPU consists of the final IR, depth, and point cloud obtained after the parallel computations have been made. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007

3. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. (a) The IR point clouds are obtained from the depth and IR images and by exploiting the intrinsic parameters of the sensor. (b) Example of six views of the resulting cloud. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007

4. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. (a) The original IR image obtained by the Kinect. It is worth noting that it appears saturated because there are many values over 255. All this values are demoted to 255, which corresponds to the white color. (b) The same image after the rescaling operation performed by the GPU. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007

5. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. Output of OpenPTrack obtained while two persons were freely moving in the scene. Each person has two trails, the green one comes from a Kinect v2 attached to a high-end laptop, while the pink one is generated from a Kinect v2 coupled with a NVidia Jetson TK1. It can be noticed that the two persons are correctly detected by both nodes. The trail generated by the embedded node is slightly less dense than the one generated by the laptop because of the differences in frame rate outlined in Tables 2 and 3. The two tracks present a small offset for visualization purposes. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007

6. Date of download: 7/9/2016 Copyright © 2016 SPIE. All rights reserved. The AAL robot based on a Turtlebot2 robotic platform, a Jetson TK1 and a Kinect v2. Figure Legend: From: Cost-efficient RGB-D smart camera for people detection and tracking J. Electron. Imaging. 2016;25(4):041007. doi:10.1117/1.JEI.25.4.041007