Object Detection Based On Ensemble Of Exemplars

Object detection is a pivotal step supporting from the low-level image process to the high-level image understanding, which is also a key issue in computer vision and pattern recognition. Object detection aims to locate the object target with a specific class detected from the static images or videos. Hence, it is a fundamental component of many important applications, such as surveillance for public safety, transportation system, intelligent navigation system and so on. However, recent methods used for object detection have great space to improve in performance. For exemplar, increasing the precision, decreasing the loss rate and enhancing the robustness and intelligence are all required to make effects on. With the improvement in performance, object detection will become more widely used in our production and living environment in the increasingly intelligent, informational and automated today, which will be of great significance to the improvement of the quality of human life and society.For the sake of increasing the precision, decreasing the loss rate and enhancing the intelligence, the article proposes a new model based on ensemble of Exemplars with the architecture of multiple instance learning, which facilitates the effectiveness of learning for the imbalance data, and the modified model works well in object detection. First of all, set the targeted class from all the samples and make the data labels, then a particular kind of HOG features are extracted for facilitating the training of classifiers with linear-svm. One classifier is trained for one exemplar from the same class. The negatives are adopted for hard mining to improve the detection precision. Meanwhile, by testing on validation set, the Co-occurrence Matrices is calculated to establish the relationship among all the classifier for a single exemplar from the same class and ensure the final model. During the procedure of object detection, a modified architecture of multiple instance learning supports to improve the performance of classifiers based on the idea of imbalanced learning, which helps to greatly reduce the noise. Besides, the classes for detection is further subdivided. For example, the proposed model can distinguish ’Bus’ or ’Car’, but not just ’Vehicle’ during detection. In the presentation of detection results, the pseudo-3D segmentation is provided for showing the more target’s information —— its shape, orientation or others, such as the orientation of a detected car, which improves the intelligence of object detection.The proposed model is established by using Matlab and C++ hybrid programming to improve efficiency. By taking the Pascal VOC2007 as the dataset, taking car, bicycle, horse as the experimental classes, and adjusting various parameters in different steps, the result of experiments proves that the proposed model outperforms the original Exemplar-SVM method, the average precision increases 21.4%,18.1%,and 26.0% for the categories of car, bicycle and horse, respectively.