Study On The Object Detection And Tracking Based On Feature Learning

As an important research topic in image processing and pattern recognition,target detection and tracking technology research has a wide range of application prospects.The application areas involve public safety,intelligent transportation,video retrieval,behavior analysis,human-computer interaction and other industries.In this paper,based on the study of traditional target detection and tracking algorithm,it is found that most of the tracking algorithms are tracked due to changes in ambient light,target attitude change and occlusion.The reason may be due to the use of descriptive features do not have good invariance or classifier selection caused by unreasonable.So,this paper focuses on how to select the appropriate feature description and create a reasonable classifier two aspects,based on the feature learning target detection and tracking algorithm to study.About object detection,this paper introduces the Adaboost learning algorithm based on the Haar-like feature description.This algorithm constructs a binary classifier,and uses the classifier to determine the target and the background.The process mainly includes the target representation and training classifier two steps.The Haar-like feature is a good indicator of the fact that the response(eigenvalue)of the different types of image blocks to the different types of Haar-like features will be different,so that the selection of the appropriate(scale and position)can be complete and accurate to express the target;training learning Adaboost algorithm using weak classifier into a strong classifier to enhance the classifier detection results.The algorithm has achieved good results in the target detection,but the robustness of the algorithm is poor because the feature does not have good invariance.In this paper,we use LBP operator to re-encode the image,and then use the normalized Haar-like feature to express the target.The new feature description has better illumination and scale invariance,so that the target detection algorithm can adapt to more application scenarios.The experimental results show that the improved algorithm is more robust and the detection effect is further improved.The target detection algorithm based on feature learning is to train the detectors by means of offline learning.This method is time-consuming and only for specific targets.It is obvious that such a detection process can not be used to the target that unknown.Considering that multi-instance online learning(MIL)can solve the problem of template drift due to prolonged unsupervised learning in the tracking algorithm,it is introduced into the tracking framework.The algorithm uses the Haar-like feature to describe the target.Based on the Online boosting learning algorithm,the classifier is selected to construct the training samples by means of online sampling.It can be found by experiments to realize most simple scenes,but the algorithm is poor in complex environment.The reason may involve that target representation and classifier learning.First,the feature description lacks invariance;secondly,the number of weak classifiers in the weak feature(weak classifier)pool is limited in order to satisfy the real-time requirement,and the original algorithm adopts the randomly selected method,and the weak classifier is selected Blindness;Furthermore,a strong classifier is obtained by linearly combining weak classifiers,which is equivalent to the contribution of each weak classifier to the final detector,which is not consistent with the actual situation.Therefore,the standard Haar-like feature and class Adaboost algorithm are used to select the classifier to achieve the target tracking.In order to verify the validity of the algorithm,a large number of data sets are used to demonstrate the experiment,and the MIL and CT algorithms are compared with the tracking effect and timeliness.The method not only can adapt to more complex scenes,but also to a certain extent,to adapt to the target small deformation,rotation and occlusion of the situation.However,there are some shortcomings in the algorithm,when the target is seriously blocked or disappeared after the algorithm failure.