Video-Based Human Motion Capture Data Retrieval

Motion Capture(MoCap)technology is a technology that uses external devices to record data and pose reduction of the displacement of human structures.In recent years,motion capture technology has been widely used in film animation production,virtual reality,mechanical control and other fields.However,acquiring motion capture data is very difficult because motion capture requires hiring actors for professional action performance,dedicated capture equipment and venues,and high human and material costs.Nowadays,motion capture technology has been developed for a long time,and people have stored a large amount of motion capture data.For the huge amount of motion capture data that already exists,the processing method of leaving it idle will inevitably cause a huge waste of resources.If we can retrieve the data that meets our needs from the already existing motion capture data,on the one hand,we can get the data more quickly and directly and improve efficiency.On the other hand,it can also avoid duplicate acquisition of motion capture data and save resources.Therefore,the retrieval of human motion capture data has important practical application value for managing and reusing motion capture data.In recent years,many motion capture data retrieval methods have been proposed,which can be broadly classified into two major categories:text tag-based methods and content-based methods.Text tag-based approaches require a lot of manual work to annotate all motion sequences in the database,which is time-consuming and text tags do not completely describe a motion sequence and add additional files that are difficult to manage.Therefore,this paper focuses on content-based human motion capture data retrieval.Content-based human motion capture data retrieval facilitates the reuse of motion data that has already been captured and stored in the database.Content-based motion capture data retrieval algorithms are broadly classified into the following categories according to the query data modality,such as motion capture data,hand-drawn sketch data,puppet motion data,Kinect skeleton motion data and video data.The query algorithm based on motion capture data is the main retrieval algorithm because the query data and the retrieved data belong to the same kind of modal data and therefore have high accuracy.However,since motion capture data is more difficult to obtain,other query algorithms for data in different modalities have emerged.Among these algorithms,the video-based query approach shows unique advantages:data acquisition is more convenient and natural,the equipment is cheap and easy to install,etc.Therefore,in this paper we propose an effective video-based method for human motion capture data retrieval.Currently,there are few video-based human motion capture data retrieval methods,and the main difficulty of this work is how to effectively describe two different modalities of 3D human motion capture data and 2D video frame sequences.In this paper,we propose a discriminative human motion descriptor based on the binary map of human contours.Specifically,for each motion capture data sequence and video sequence,we first compute its corresponding human contour bipartite map.For the human motion capture data sequences,we obtain the human binary contour image for each motion capture data frame by rendering the human model projected to a specific view direction.For video data,we obtain the human body binary contour image for each video frame by background differencing.Further,we put the extracted set of human contour images as input into the convolutional neural network MotionSet to extract its corresponding representative motion feature vectors to obtain valid motion descriptors.Finally,we perform the retrieval of human motion capture data by performing similarity measures on the extracted valid motion descriptors.The focus of this research is to propose a novel algorithm for cross-modal human motion capture data retrieval.The core part of the algorithm is to extract discriminative cross-modal motion features,which is implemented by converting the original video or motion capture clips into binary map sequences of human motion contours,extracting their representative motion features through the MotionSet network,and then calculating the similarity using the nearest neighbor approach to obtain the final retrieval results.Experiments show that our proposed algorithm achieves an improvement of about 0.25 in average MAP compared with the benchmark algorithm.