Research On Object Detection Based On Two-Stage Boosting Oriented Gradient Feature

Object detection is the process of fixing a position from complex background and extracting foreground objects of interest. It has broad application prospects in intelligent control system, human-computer interaction, image retrieval based on content, number plate recognition and many other fields. In addition, object detection is still the foundation and premise of advanced treatment of various follow-up, for example, incident detection and object classification, behavior recognition, behavior analysis and semantic indexing and so on. Therefore, research on object detection has its theoretical significance.In this paper, we analyze important document of object detection in recent years. Up to now, object detection has been several decades of development history. According to different information of use, object detection can be divided into two categories in general, object detection based on motion information and object detection based on feature information. Further, object detection based on feature can be subdivided into methods based on object model, method based on template matching and method based on statistical classification methods. And from all the current detection method, since object detection based on statistical classification methods has a good robustness, and it also is one of the best performance algorithms currently, and then it has become mainstream in the field of object detection method. Object detection based on statistical classification method has two key steps: feature extraction and classifier design. The quality of these two aspects is affecting the merits of the ultimate detector performance. Classifier design actually is the choice of machine learning algorithms. According to the different choice of machine learning algorithms, object detection based on statistical classification method can be divided into methods based on neural networks, methods based on support vector machine and methods based on Adaboost. The main work of this paper is that we improved part on featured extract which is part of object detection method based on Adaboost, and proposed object detection method based on Two-Stage Boosting Oriented Gradient Feature.Object detection based on statistical classification method is that learn and gain a classifier from a range of training data by machine learning, then use the classifier to identify the input window. In this paper, we particularly in-depth study of two most typical feature-based detection algorithm: object detection based on Hog+SVM proposed by Dalal and object detection based on feature-Shapelet proposed by Sabzmeydani. Their detection performance is achieved very good results, but since feature vector size of Hog is too large, and then to cause computational overhead, and the shortcomings of long training time. Focus on this problem, this paper improved parts of feature extraction with Hog for the prototype, and proposed Two-Stage Boosting Oriented Gradient Feature-TSBOG feature.The method of TSBOG feature extraction can be expressed as: First, the input image will be divided into several blocks, which compose of four cells shaped by the field structure. This blocks divided from input image may overlap or not, in case of overlapping detection performance is more accurate. Second, count the distribution of gradient of all pixels in all direction for each cell. In general, the range between 0 to 180 degrees direction can be divided equally into 9 bins, one bin for every 20 degrees, 0~180 degrees and 180~360 degrees can be classified by method of equal to the isometric, and then can derive feature vectors of cells, each cell of a 9-dimensional feature vector. Then we connect those feature vectors of four cells in each block, and gain a block of 36-dimensional feature vector. Third, input blocks of 36-dimensional vector into the Adaboost machine learning algorithm, to select some of particular components of gradient direction which have excellent classification ability, and then construct TSBOG feature which has a self-adaptive characteristics. The biggest advantage of TSBOG feature is the self-adaptive characteristics. Compared with the other underlying feature which is fixed, TSBOG feature which is upgraded from machine learning algorithm can obtain more information for classification, while some redundant information may weeded out. So improve the detection efficiency of feature, and improve the robustness of object detection.This paper proposed object detection based on Two-Stage Boosting Oriented Gradient feature. The core idea of this algorithm is twice Adaboost: First, focus on blocks from local image, all low-level feature in blocks of local area of image train by first Adaboosst, and after training can gain oriented gradient feature, Hog feature; Then, all TSBOG feature which generated from first training process will be inputted to have second training by Adaboost, and choose the best combination of weak classifiers to construct the final strong classifier. In addition, in this paper many various parts of algorithm process have been optimized, including feature extraction, feature pre-selection and weak classifier and so on.First, to optimize the speed of detection, in the phase of feature extraction, this paper use integral image with histogram to speed up feature vector calculation. And then while we calculate feature vector of one cell, each quantization interval only need to read memory four times and carry out the addition and subtraction operation. This method is not affected by the size of feature vector, and greatly improve the speed of detection.Second, this paper proposed a formula which can judge the merits of feature. Take advantage of overlapping degree of distribution density to judge the cohesion of feature, and take a view of classification ability of feature from essential, then can evaluate the merits of a feature. Compared with traditional method which judge merits of a feature by adding error rate, this method proposed by paper have better promotion. Finally, this judgement criteria is applied to process of pre-selection, optimize the feature number and training process, greatly reducing the system training time.Third, this paper also improves the structure of weak classifier. We use real-Adaboost and Look-up table instead of Stump in classifier training. The main disadvantage of the threshold classifier is that it is too simple to fit complex distributions. While Look-up table classifier can output a real-value which can represent confidence, it can optimize the structure of classifier, and greatly improve the performance of object detection.The experimental results show that object detection based on Two-Stage Boosting Oriented Gradient feature in this paper has improved the detection rate, lower false alarm rate in a large extent, and the confidence level of target recognition has been markedly improved, and the algorithm presented in this paper can get better performance of pedestrian detection in various natural background .