ABSTRACTWe propose a new approach which is a three-stage pipeline to fast and accurate segment hand from a single depth image. Firstly, a depth frame is segmented into several regions by histogram-based thresholds selection algorithm and tracing the exterior boundaries of objects. We found that MINIMUM, MEAN and MEDIAN are effective ways to separate objects and the threshold in the valley between two maxima similar to MINIMUM algorithm with a minimum error. Then, each segmentation proposal is evaluated by a 3-layers shallow convolutional neural network (CNN) which is trained as a binary classification function to predict whether it is a partition of hand. Finally, all hand components are merged as our hand segmentation result. In our experiment, we use a set of real data containing more than 200,000 frames of depth images. Compared with the results achieved by approaches based on RDF and SegNet, results demonstrate that our approach achieves better performance in high-accuracy (88.34% mean IoU) within shorter processing time (8 ms).
Read full abstract