Research And Implementation Of Pedestrian Detection Algorithm Based On Feature Fusion And Online Learning

Pedestrian Detection is a classic research project in the area of computer vision. At the same time, it is also the vehicle driver assistance, intelligent video surveillance and other key technologies of computer vision applications, with a high value. Pedestrian detection method is mainly divided into pedestrian detection based on background modeling and foreground modeling.The former is to model static backgoround and extract dynamic foreground from the background image.The latter is to model foreground and learn pedestrian features from the training sample so that establishing the pedestrian model, the method of detection generally use the sliding window scanning strategy that searching exhaustively positions of all pedestrians emerging from the image. Because pedestrian detection method based on background modeling exists an implied prerequisite that having an almostly static backgorund, a big movement in the image can be considered foreground, when the target is stationary, or camera movement,the detection method often fail, the current hotspot for pedestrian detection is mostly focused on the foreground modeling approach.In fact, due to the diversity of pedestrians appearance and background illumination changes and other factors, it is very difficult to effectively complete a long, highly accurate pedestrian detection task for using a single fixed template of the foreground. According to the above problems, this paper mainly focuses on studying the pedestrian detection algorithm based on feature fusion and online learning,and the research content of the specific as follows:(1)In feature extraction, this paper adopts the strategy of mixing different features, using integral channel characteristic way to introduce a variety of features to pedestrian detection method,and on the basis of analysis and testing a variety of single feature, combining organicly multiple features. The second chapter introduces the various features performance test,and determine the most appropriate combination of features to detect pedestrians.(2)As to the pedestrian learning algorithm, this paper selects online Adaboost algorithm as an online feature learning algorithm. Unlike traditional offline pedestrian learning algorithm, on the basis of the existing test results, it can keep on learning the pedestrian features in the scene so that the pedestrian detection algorithm has strong adaptive and self-learning ability.(3) Aiming at difficulties of the automatic tagging samples category in online learning, this article uses collaborative training algorithm for automatic tagging test results obtained by each frame, at the same time, in view of the Tri- training algorithm trained classifier problem with low independence was proposed based on improved algorithm under the framework of Tri- training, using PCA classified sample set to make it meet the two redundant view and sufficient conditions, structural differences enough three datasets, training the talented and more independent classifier. By using Tri- training joint training strategy, this algorithm can use effectively no label samples, and implement a semi-supervised learning, a large number of experimental results show the effectiveness of the method.In the end, this paper summarizes the work done, and analyzes and discusses the further research plan for the shortcomings of this paper.