Understanding the movement plays an important role in the analysis between the media based on video and learning many knowledge. A group of researchers led by a fan of Hehe studied problems with recognizing and predicting physical movement using deep neural networks (DNN), in particular weave neural networks and recurrent neural networks. Scientists have developed and tested the approach of deep learning based on a relative change in the coded position as a series of vectors and discovered that their method exceeded the existing traffic modeling framework.
In physics, movement is a relative change in position in time. To eliminate object -oriented factors and backgrounds, scientists focused on an ideal scenario in which the dot moves in a two -dimensional (2D) plane. Two traffic modeling tasks were used to assess the ability of DNN architecture: traffic recognition and predict movement. As a result, a vector network (VECNET) was developed for modeling a relative position change. The key innovation of scientists was the coding of the movement separately from the position.
The group's research was published in the journal Intelligent calculations.
The study focuses on movement analysis. Traffic recognition is aimed at recognizing various types of movements based on a series of observations. This can be seen as one of the necessary conditions for recognizing action, because recognition of activities can be divided into recognition of objects and traffic recognition. For example, to recognize the “Open door” campaign, DNNS must recognize the “door” object and “open” movement. Otherwise, the model would not distinguish the “open door” from “Open the window” or “open the door” from “Close the door”. Movement forecasting is aimed at predicting future position changes after seeing a part of the movement, i.e. the context of movement that can be considered one of the required conditions of video forecasts.
Vecnet takes a short -range movement as a vector. VECNET can also transfer DOT to the appropriate position given by the vector representation. To get access to motion with time, for aggregation or prediction of the vectors representation (LSTM), a long short -term memory (LSTM) was used. A new VECNET+LSTM method has been created can effectively support both recognition and anticipation, proving that modeling a relative position change is necessary to recognize traffic and makes it easier to predict movement.
Recognition of activities is related to traffic recognition because it is related to traffic. Since there is no clear DNN architecture to recognize activities, scientists compared and examined the subset of models covering most of the domain.
The VECNET + LSTM approach has gained higher results in traffic recognition tests than six other popular DNN architectures from video tests regarding the modeling of relative position change. Some of them turned out to be simply weaker, and some were completely inappropriate to model traffic.
For example, compared to the Convlstm method, the new method was more accurate, required less training time and did not lose its precision so quickly when making additional forecasts.
Experiments have shown that the VECNET + LSTM method is effective in recognizing movement and prediction. It confirms that the use of a relative position change significantly improves traffic modeling. In the case of methods of processing appearance or image, you can use the offered method of modeling movement for a general video understanding that can be studied in the future.
