Aiming at the problem that the text results of continuous sign language recognition based on video have problems of semantic ambiguity and chaotic word order, a two-step method was proposed to convert the sign language text of the continuous sign language recognition result into a fluent and understandable Chinese text. In the first step, the natural sign language rules and N-gram language model ( N-gram) were used to perform the text ordering of the continuous sign language recognition results. In the second step, a Bidirectional Long-Term Short-Term Memory (Bi-LSTM) network model was trained by using the Chinese universal quantifier dataset to solve the quantifier-free problem of the sign language grammar, so as to improve the fluency of texts. The absolute accuracy and the proportion of the longest correct subsequences were adopted as the evaluation indexes of text ordering. Experimental results showed that the text ordering results of the proposed method had the absolute accuracy of 77.06%, the proportion of the longest correct subsequences of 86.55%, and the accuracy of quantifier completion of 97.23%. The proposed method can effectively improve the smoothness and intelligibility of text results of continuous sign language recognition. It has been successfully applied to the video-based continuous sign language recognition, which improves the barrier-free communication experience between the hearing-impaired and the normal-hearing people.

In view of the problem that conventional incremental learning models may go through a way of performance degradation during the update stage, a kernelized incremental learning method was proposed based on objective equilibrium measurement. By setting the optimization term of "empirical risk minimization", an optimization objective function fulfilling the equilibrium measurement with respect to training data size was designed. The optimal solution was given under the condition of incremental learning training, and a lightweight incremental learning classification model was finally constructed based on the effective selection strategy of new data. Experimental results on a publicly available fall detection dataset show that, when the recognition accuracy of representative methods falls below 60%, the proposed method can still maintain the recognition accuracy more than 95%, while the computational consumption of the model update is only 3 milliseconds. In conclusion, the proposed method contributes to achieving a stable growth of recognition performance as well as efficiently decreasing the time consumptions, which can effectively realize wearable devices based intellectual applications in the cloud service platform.