Air-writing characters modelling and recognition on modified CHMM
Songbin Xu, Xue Mei Yang · 2016
In this paper, we provide two databases DB1 and DB2 of motion characters written in the air, and present an accelerometers and gyroscopes based air-writing characters recognition system. The DB1 of 10 characters was collected by 40 subjects without writing constraints, while DB2 of 36 characters was collected by 49 participants in constrained stroke orders. We preprocessed the raw data with Moving Average filter and Z-score normalization, then utilized a modified CHMM for motion characters modeling and the Viterbi algorithm for recognition. We utilized an ASD rule for states assignment and managed to avoid the underflow issue during training. We analyzed the effect from limitations in writing. Results show that our system can realize real-time use and achieve high accuracy in both the mixed-user and the user-independent style.