Research On Action Recognition And Its Application Based On Depth Images

Human action recognition is an important yet challenging task for computer vision.Recent years,human action recognition has drawn much attention due to its applications in various domains,such as visual surveillance,human-computer interaction(HCI),virtual reality.Relying on RGB pixels,traditional models of human action recognition suffer from the distraction of background and lighting,which can hardly be applied in engineering.So,recently developed commercial depth sensors brings new measures to tackle this task.This dissertation concentrates its attention on the research of human action recognition using depth images,and the main contents and contributions of this dissertation are as follows:1.In order to identify the sequential orders of Depth Motion Maps and improve the efficiency of feature extraction,a new feature extraction method using adaptive sampling pyramid model is proposed.This model partitions frames on basis of motion energy,which constructs multi-scale features to describe the features in different scales,and improves the model robustness for temporal variance.Applying the improved sparse representation based classification algorithm,experimental results on the military sign language action dataset show that this approach can achieve the accuracy of 96.875 percent when the observation is placed in the location of C0.2.Constructing a military sign language action dataset.According to the characteristics of military sign language action,eight representative actions are designed for the dataset construction,collecting depth images and skeleton points from five observations innovatively.This dataset contains multiple depth images and skeleton points describing human actions,which meets the objectives of algorithm developing and testing.3.A human action recognition algorithm based on depth image multi-feature fusion,analyzing human action observed from different views.After analyzing the model of multi-sensor image fusion,a depth image fusion model is constructed and Fisher criterion based algorithm is applied to realize the recognition under multiple observation locations.The algorithm choses the optional discriminant vector by using the principle of minimizing the sum of the intra-class distance,which improves the average accuracy of human action recognition under multiple observation locations from 93.07 percent to 98.75 percent.4.In order to solve the problem that most human action recognition algorithm cannot deal with continuous depth image sequence,a new method based on skeleton joint is proposed to detect the starting frame and partition frames.The method can judge the starting frame and extract useful frames in real time by the calculation of skeleton coordinate values.The proposed method is efficient and effective,which meets the need of practical application.5.In order to solve the problem that the human action recognition algorithm that requires large memory when dealing with the continuous depth image sequence,an improved algorithm is proposed.On the basis of a detailed analysis of the memory allocation mechanism of MATLAB,the algorithm structure is optimized and the data processing flow is changed to solve the problem.The above-mentioned research work of this dissertation provides fundamental theory and technique support to the establishment of the military sign language action recognition system based on human action recognition algorithm using depth images,which is of great practical and strategic significance to the improvement of information warfare as well as the establishment of the unmanned combat platform.