Intelligent Fuzzy Network for Dynamic Sign Words Recognition from Spatial Features

Dynamic sign words recognition (DSR) from Spatial features using skeletal hand joints information play an important role in similarity problem of sign words. Unfortunately, recognizing dynamic sign words from spatial features are challenging as movement features have variety of complex appearances. Existing methods do not capture spatial interactions and dependencies among the estimates, which are extremely significant for DSR. To solve this problem, since spatial frames contain non-periodic and fast transient features, we decompose features into 3D wavelet transform (3D-WT) signals. 3D-WT maps each frame into other basis functions, to encode spatial relationship. 3D-WT features are fuzzified and utilized as input vector in Adaptive-neuro fuzzy inference system (ANFIS), which adaptively learn 3D spatial characteristic of hand joints. Then, ANFIS match and fuse features by iterative learning, which learns rich contextual information to recognize dynamic sign words. ANFIS achieved recognition accuracy of 94.08% and 93.16% for single and double-hand words. Extensive experiments by proposed approach on public data sets outperform some existing methods.