Research On Dynamic Object Tracking And Related Applications Based On Deep Learning

In the future Internet of Things era,sensor devices will be everywhere in our social life,and they will continuously collect our physiological data and motion data.Therefore,based on the inertial sensor and deep learning model,this paper proposes a precise reconstruction algorithm for trajectory and an algorithm for action segmentation and recognition,and proposes an adaptive skewness kurtosis structure that can improve the feature integration capabilities of deep learning models.Because inertial sensors have large measurement errors and the theory of inertial guidance is very sensitive to sensor errors,it is hard to use them to complete accurate object trajectory restoration and tracking tasks.But compared with optical sensors,inertial sensors are small in size,easy to wear,and can independently complete data collection.Therefore,we designed a corresponding method to complete the task of accurate trajectory restoration based on inertial sensors.We take hand movement as the main research object,and adaptively divide the hand movement trajectory into a combination of several basic trajectories by designing a suitable deep learning model.We also set up a geometric model library,which can complete high-precision restoration of various basic trajectories with the help of deep learning models.Coupled with the corresponding interpolation smoothing process,the high-precision hand motion trajectory restoration is completed.This method decomposes the restoration of complex trajectories into the restoration of several basic trajectories,and converts the restoration of basic trajectories into the prediction of geometric model parameters,and finally makes up for the lack of hardware with the advantages of the algorithm.In order to achieve human-computer interaction on the basis of trajectory reconstruction,we need to accurately extract sensor data corresponding to specific actions from the dynamic data stream.Therefore,we combined the theory of inertial guidance,object detection algorithm,Fréchet distance and deep learning model to establish a specific hand movement dynamic recognition algorithm.This algorithm is not only effective for inertial sensors,but also has a very good dynamic recognition effect for optical sensors.The data is converted into two-dimensional or threedimensional geometric trajectory data,and an enlarged prediction window is given through the LSTM model we designed.After calculating the Fréchet distances between the two spline functions we set and the trajectory in the window,two similarity curves can be obtained,and finally set the corresponding recognition algorithm to get the precise start and end points of the action.Since the inertial guidance method is more sensitive to sensor data,our method is more effective in identifying inertial sensors.For optical and inertial sensors,this method can control the dynamic recognition accuracy of specific actions within 10 and 5 frames,respectively.This paper studies the motion tracking problem based on the deep learning model.Therefore,in order to improve the feature integration capabilities of deep learning models,we design an adaptive skewness-kurtosis neural network(ASKNN),which can adaptively modify its skewness and kurtosis without changing the mean value and variance of the neural network layer,and at the same time allows the neurons in the layer to carry out effective information communication and exchange without significantly changing its statistical properties.This ensures that the ASK structure will not greatly change the characteristics of the original deep learning model,and can well retain its advantages.In this way,the ASK structure enhances the feature integration capability of the deep learning model,which is particularly effective in the face of noisy data.