Research And Implementation Of Parallelized Human Behavior Recognition Method

Human behavior recognition is currently a hot research field, which involves a lot of research areas, such as sensor system, machine learning, artificial intelligence and computer vision. With the emergence and popularity of somatosensory equipment, the use of somatosensory equipment for human behavior recognition is becoming more popular. As an outstanding representative of them, K inect has excellent depth perception of scene ablility and three-dimensional body joints recognition ability; it is largely used in the virtual reality, intelligent home field, sports analysis, etc.Meanwhile, with the sharp increase in the amount of data and the size of data, the traditional identification algorithm has been unable to meet the recognition needs, so the parallelization of traditional behavior recognition algorithm is imperative. Spark is currently the most widely used general-purpose parallel computing framework, whose resilient distributed dataset model greatly improves the efficiency of distributed computing and data-intensive computing. The study of human behavior recognition algorithms on Spark framework has a great value and significance.Human behavior recognition has two parts, the first is behavioral models feature extraction and the second is classification model building. In the feature extraction of human behavior, this article takes the human skeletal structure characteristic vector and angles from the static behavior, and they are also coupled with the skeleton factors. For dynamic behavioral characteristics, this article takes the key frame body structure similarity from the self-adapted search algorithm. The static characteristic sequence acts as a dynamic characteristic expression, and the threshold of search algorithm is determinated by experiments in real conditions; the results show that the extraction of features is discriminative.In behavior recognition algorithms, this paper focuses on the behavior of classification methods which is based on machine learning, a parallel neural network algorithm called PANN based on Spark is designed and implemented. On the training phase, the algorithm parallelizes data to subtasks and broadcasts the weight matrix to clusters as a global variable. Massive iterations on large datasets show significant efficiency gains to the traditional algorithm. At the same time, the L-BFGS algorithm is substituted for the gradient descent algorithm in BP neural network, which optimized the performance of behavior recognition.Based on the behavior feature extraction method and the parallel behavior recognition algorithms, this article implements a complete daily behavior recognition platform. At the input layer, the behavioral data is collected from the K inect system; the HDFS distributed file system is the layer of data storage, the upper Spark platfor m level acts as a feature extraction and parallel execution environment. A large number of experiments have conducted to prove the platform’s usefulness, effectiveness, and scalability.Results of this study proved the feasibility of parallelized human behavior recognition method, and the methodology can be used in many scenarios.