Increasing the Robustness of CNN-Based Human Body Segmentation in Range Images by Modeling Sensor-Specific Artifacts

This paper addresses the problem of human body parts segmentation in range images acquired using a structured-light imaging system. We propose a solution based on a fully convolutional neural network trained on realistic synthetic data that were simulated in a way that closely emulates our structured-light imaging system with its inherent artifacts such as occlusions, noise and missing data. The results on synthetic test data demonstrate quantitatively the performance of our method in identifying 33 body parts, with negligible confusion between the front and back sides of the body and between the left and right limbs. Our experiments highlight the importance of sensor-specific data augmentation in the training set to improve the robustness of the segmentation. Most importantly, when applied to range data actually acquired by our system, the method was capable of accurately segmenting the different body parts with inter-frame consistency in real-time.