Human Action Interpretation Using A Compact DTA Framework

Electronic devices have achieved unique functions due to the emergence of action recognition using deep learning methods.It has gained tremendous recognition from computer vision researchers due to its tremendous potential in the field of human computer interaction and visual analyzes.However,there are various challenges related to the computational cost and efficiency of this approach.Human action recognition is a sophisticated field in computer vision,because an image sequence consists of various features that do not belong to a specific action.To address these issues,a novel approach is proposed for human action recognition and localization based on DTA terminology an abbreviation of detect,track and analyze.An innovative architecture is presented which is inspired by the state-of-the-art methods including yolov3,deep-sort and 3D convolutional neural network.The yolov3 is employed with the darknet-53 architecture to predict the bounding boxes of people from an input image,and then followed by the deep-sort algorithm to track the detected people.Finally,region of interest(ROI)version of KTH dataset is created to train the proposed 3D convolutional neural model to recognize human action from the tracking sequence.In addition,we extensively discussed the traditional and modern methods of object detection,tracking and human action recognition.To thoroughly test the proposed approach,we have performed various experiments.A proper method is presented to analyze human action by developing the DTA framework using a mono camera.The action recognition model is trained and tested using the videos of six human actions,in which people are cropped from the background,and then the tracked image sequence of them are given to the trained model for predicting the correlated action.The ROI version of the KTH dataset has been created to train and evaluate the human action recognition model.The experimental results indicated that the accuracy and the estimated error of the proposed approach are better compared to other methods.The overall framework is very efficient and straightforward that most of the modern devices are able to adapt this method for visual analyzes with tracking and localization capabilities.Overall,the results showed that the proposed method outperforms previous state-of-the-art methods by various aspects.The visual analyzes system is able to run between 10 and 13 FPS,while the action recognition model has achieved the accuracy of 97.31%.