As Google's artificial intelligence AlphaGo and Li Shishi's five rounds of man-machine war ended, the Alpha Dog eventually took a big 4:1 battle against Li Shishi, and people's curiosity about the Alpha Dog was pushed to a climax. Alpha Dog also showed the powerful ability of deep learning neural network (DNN) to people.
What is DNN in the end? DNN is a hot topic in the field of artificial intelligence in recent years. Its applications in speech recognition, automatic driving, and handwriting recognition are extremely successful. In the field of artificial intelligence, neural network is a model for simulating biological nervous system. It is composed of many unidirectionally connected neurons and can convert input signals into output signals according to various parameters of linked neurons. Compared to ordinary neural networks, deep neural networks (DNNs) use implicit multi-layered complex structures and nonlinear transforms to express a high degree of abstraction of data. These features are closer to the human brain and it is easier to achieve certain functions of the human brain, such as the brain functions needed to play chess.
Alpha Dog has applied two kinds of deep learning neural networks, namely "strategic network" and "value network" to fit the situation strategy function and evaluation function. It can be said that this is the two brains of Alpha Dog. The victory of the Alpha Dog detonated the concept of DNN. In today's world of big data, deep learning neural networks have been applied to many fields and become an important tool for revealing scientific principles and upgrading existing business models.
So can the cool DNN be used in the hot VR industry? The answer is yes. At present, in the field of VR applied to DNN technology, including speech recognition, gesture recognition, etc., the author here mainly introduces the DNN in the gesture recognition algorithm.
There are two kinds of traditional gesture recognition methods: gesture recognition method based on data gloves, gesture recognition method based on computer vision.
The computer vision-based hand gesture recognition method has become a research hotspot because it does not depend on the equipment, the more natural human-computer interaction effect, and the better immersiveness. Based on the computer vision gesture recognition, the gesture image information is obtained from the camera. After appropriate data preprocessing, the gesture is segmented from the image. After the feature extraction is performed on the segmented gesture, the gesture template is used for classification.
The traditional methods of classification are template matching, fingertip detection methods, etc. However, these methods need to manually extract the target features from the image and write templates to match. As deep learning neural network is proposed, gesture recognition based on deep learning neural network (DNN) also comes into being. This gesture recognition algorithm has a great advantage in all aspects compared to the traditional gesture recognition algorithm.
The working method of deep learning neural network comes from the analogy of the human brain's visual mechanism. This process of starting from the original input to the non-stop abstraction of the higher level gives the model a high degree of abstraction, making the deep learning neural network very effective. Deeply extracting data feature information from a large number of tagged data, fully mining data intrinsic attributes and valuable characterization data, and then combining low-level features into more abstract high-level features, while advanced features are more advanced and more essential data. Description, which can get better results on the classification problem.
As we all know, the VR industry is facing a major technical problem is the processing of massive data, especially for gesture recognition technology, the hand has a very large number of joints, requires a very strong recognition ability to accurately identify each fine movement. The multi-hidden layer structure of the deep learning model enables the model to effectively use massive data for training. The more data used, the higher the performance of the model, making it ideal for gesture recognition in VR environments.
In computer gesture-based gesture recognition, traditional algorithms cannot directly and effectively extract useful information from the image. However, the learning ability of deep learning is extremely powerful. Even complex low-resolution images can extract target depth features very well. The image background needed for gesture recognition based on DNN does not need to be fixed, and the algorithm allows motion background even within a certain range, thereby improving the recognition of environmental tolerance and fineness.
| How does DNN achieve in gesture recognition?Next, the author will introduce the specific implementation method.
The traditional identification system process can be roughly divided into three steps: detection, identification, and tracking. The specific implementation process is: sensor acquisition information, preprocessing, feature extraction, feature selection, and finally to the final reasoning, prediction or identification. It is generally believed that the last part belongs to the part of machine learning. This part is the essence of the whole system. Whether it can learn effective knowledge from the data is directly related to whether the entire system can work as people expect, but it also has the characteristics of the previous period. Extract or choose a bad situation, this will affect system performance.
The DNN-based gesture recognition is to allow the machine to extract features, do not need manual feature extraction, and powerful learning capabilities make the model can achieve the desired effect in a complex context. The gesture recognition process is as follows:
Firstly, a gesture recognition acquisition system is created to detect the moving target, and at the same time, part of the information of the foreground, that is, the moving target is extracted and the gesture image is obtained. After acquiring the gesture image, an appropriate network protocol is selected according to different application scenarios, and the data is transmitted to the computing platform. The deep learning algorithm is used to perform complex background gesture recognition, and finally the gesture recognition result is given.
| What is VR based on DNN?A VR product based on the idea of ​​DNN algorithm, and uses a stand-alone GPU method to accelerate the training and recognition of deep networks. Gesture recognition module can use the depth camera to achieve close-range 3D imaging of the hand, combined with DNN independently developed a set of data processing algorithms to achieve high-precision real-time hand movement recognition, can not only track a single finger, identify each finger subtle movements Can also be extended to track multiple hands.
The gesture recognition of such a VR product has high capturing accuracy, fast response, high sensitivity, and can be used both indoors and outdoors without being affected by the ambient light intensity.
However, there are almost no DNN-based gesture recognition products on the market. The StepVR product developed by G-Wearables, which is known by the author, is based on the DNN algorithm, but it hasn't been seen yet. We can look forward to it.
Shenzhen ChengRong Technology Co.,Ltd. , https://www.dglaptopstandsupplier.com