Study On An Object Recognition Algorithm Based On HOG

Object recognition is an import research field in computer vision and it enables acomputer to simulate human thinking process to identify a specific object activities. Nowit has been widely used in the industrial production testing, quality control, biomedicineimage enhancement, intelligent traffic video monitoring, remote sensing image analysis,military image scanning and other fields. A good object recognition system can analyzeand manipulate the camera image sequences without manual intervention, so that it canidentify specific goals and track them better, meanwhile improve recognition accuracy andefficiency.An object recognition system based on the HOG features use sliding windowmechanism to extract the appearance of the image edges and locate them. It mainlyincludes two steps: feature extraction and test. First it extract HOG features of sample dataand represent them in the feature vector, then it use the appropriate classifier to train them,finally complete the object detection and recognition tasks. It uses the technology of edgedetection, HOG feature extraction, classifier training algorithm. First of all it normalizethe color space of images to reduce the impact of illuminations. Using a symmetric centertemplate to extract edge gradient direction features is the best when we extract edgefeatures. In order to use the local gradient edge information effectively, we use slidingwindow mechanism to scan the image densely and normalize the overlapped block. Thesize of cell and block is important to the result. Finally we need to gather all of the HOGfeatures of target window and use a vector to represent the feature of the whole window.Then we can use support vector machine to train them two times in order to improve theaccuracy. We can make a number of weak classifiers with different weights to combine apower strong classifier. At each level we use AdaBoost Algorithm to get a strong classifierand connect them together to form a cascade classifier. Its error rate is vey low, buttraining time is longer. In order to reduce computation time, we can use decision tree toimprove it.Based on the analysis of detection results of different objects, object recognition algorithm based on the HOG features can locate the object under complex background andmeet the need of real time and accuracy. It can be used for video monitoring andintelligent traffic of cities.