Segmentation And Recognition Model For Continuous Action Sequence

Human activity recognition based on inertial sensors has always been a research hotspot.In recent years,with the continuous development of computer technology,microelectronics technology and integrated circuits,sensors have made extraordinary breakthroughs in volume,computing performance,and manufacturing costs.Therefore,human movement collection and recognition based on smart phones and wearable devices are widely used and play an important role in health detection,games and entertainment,medical rehabilitation,robotics and other fields.Most human activity recognition focuses on the classification and recognition of a single behavior or a single body movement.However,in actual human activities,multiple motion states alternately exist,and the monitored signals are often continuous and complex time series.In the face of real complex actions,it is difficult to accurately segment and recognize the existing action information and motion state information.In view of the above problems,the continuous motion segmentation and recognition based on inertial sensors have been explored and studied.The main work of this paper includes the following aspects:1.Divide basic human daily actions into three categories:dynamic actions,static actions and transition actions.Among them,the dynamic actions mainly include walking,upstairs and downstairs.This series of actions are mainly based on human walking of the lower limbs.Static actions are mainly collected when the human body is still,including sitting,standing,and lying down.Transition actions are actions that transition from one type of basic action to another type of basic action.For example,there is a sit-to-stand action transition between sitting and standing actions.2.Based on the sliding window segmentation method,the dynamic action and static action information in the continuous action sequence are obtained.The detection and segmentation of static action is based on the characteristic that the rate of change of the static action signal is close to zero,and the static action signal fluctuation threshold is given by the statistical method,and the static action area is further subdivided,and then the transition action can be changed according to the action state conversion relationship and the action a priori Precise segmentation.The sliding window segmentation method is not accurate enough to segment the gait motion.The gait signal is detected according to the periodicity of the gait motion signal,and the starting position of the complete gait signal is determined by detecting the local minimum,so the signal can be extracted Signals with such rules are used as candidate segments for dynamic signals.The candidate segment can obtain the action category through the classifier.3.Finally,for all the accurately segmented action segments,use the trained LightGBM classification model to classify and determine the action information of each segment.In the collected data set,the overall segmentation accuracy and recognition rate can reach more than 98%,which can well extract various action information and regions in the unknown complex continuous action sequence.