Research On Object Detection And Action Recognition Based On Random Forests Algorithm

With the development of science and technology innovation capability in the world, the traditional video surveillance technology with non-modern label can not meet the needs of people. How to improve the monitoring system intelligently has become an important research topic. This involves a lot of image processing research techniques, for example, real time capture of moving object in video image, setting up an analytical model to understand the content of the scene in the monitoring system and more advanced automatic alerts for abnormal human behavior. These studies have prompted more scholars to devote themselves to the field of computer vision, which has made great contributions to the modernization of industrial information. Focusing on existing massive video data has great research value, this paper proposes a novel algorithm based on moving object detection and character recognition by using the machine learning technique of random forest. With the different of methods that stay in the image level about pixel processing, this paper empowers the model’s ability of classifying the object discriminately, by excavating deeply the image feature information, supervised learning framework and a variety of image processing technologies.In the aspect of object detection, the feature information is particularly important for the representation of the object. On the basis of previous research, this paper supplements the feature vector of histogram of oriented gradient by introducing Haar wavelet tools, which aims to extract the high and low frequency signal of image. Haar oriented gradient descriptor is formed by the signal, which is connected in series via the algorithm of histogram of oriented gradient. Then, based on the feature information provided by the patch, the generation of incremental random tree is redefined. The voting image is acquired via the node. The extreme value of probability density is obtained by applying the Meanshift algorithm. The region is split and extracted effectively which contains the object information. Finally, two rules are introduced to evaluate the classification effect of the classifier and the detection accuracy of the image. The accuracy of the algorithm is verified through 7 sets of experiments.In the aspect of action recognition, this paper summarizes three aspects, which are the presentation layer of feature information, the layer of construction model and the layer of action classification. In the presentation layer of feature information, For this problem that optical flow is hard to describe character movement information, based on original brightness constant assumption, this paper adds hypothetical condition, including gray intensity invariant in the gradient direction, velocity component field smoothing, the continuity of regional descriptor matching and the consistency of adjacent frame descriptors. The computation of the optical flow field is coded via color information, which aims to show the movement information of the object. Experimental results show that the optical flow algorithm can describe the change information of the moving objects, and provide the necessary content for the following two layers. In the layer of action classification, in view of the fact that action has coherent information in time-space domain, this paper proposes a model construction method based on patches. The random tree is redesigned by many methods, including increasing the time information in the original 2 dimensional image, combining with the criteria of the identification of the node and forming the incremental growth. In the layer of action classification, action classes are determined rely on the action map that is generated via the node. In this paper, we have carried out experiments on 3 sets of databases, and the accuracy of the algorithm is verified by the comparison of the confusion matrix and the action recognition accuracy.