Research On Pedestrian Detection Based On Classification For Changing Scenes

Pedestrian Detection Systems (PDS) aim to indentify and localize pedestrian in front of a moving vehicle accurately and quickly, with the help of intelligent processing algorithms through some sensors which installed in the car; and then to alert the driver and to make some emergency control if necessary. The purpose of PDS research is to ensure pedestrian life and property. With the development of intelligent vehicles, pedestrian detection systems achieve more attention by industry as a core enabling technology. Meanwhile, PDS research involves sensors, machine learning, automation, information integration and computational intelligence. It is also a hot research topic in the cross-discipline flied such as intelligent monitoring, which has great theoretical research value and real application value.Comparing to general static detection systems, pedestrian detection system is a dynamic detection platform. The diversity and time-varying of traffic environment and the dual role of the dynamic detection platform and detection object, making the pedestrian detection technology face a series of technical difficulties. Pedestrian detection in changing scenes becomes an acknowledged challenging problem. It need to design a highly efficient detector and can adapt to different scenes.Currently, pedestrian detection techniques in changing scenes can divide into two categories including image processing and machine learning. The first category focuses on using image processing technology to process the pedestrian image, and further detect an object as a pedestrian. Pedestrian detection method based image processing can not meet the requirement of real-time pedestrian detection because of complex background of pedestrian detection systems. Classification methods in machine learning are used more as a pedestrian detection method, and it has achieved successful applications in the fixed scenes. However, the scenes of pedestrian detection are dynamic and changing. The diversity and dynamic of detection scenes make the traditional classification methods often can not meet the requirements of pedestrian detection in changing scenes. This paper focuses on pedestrian detection method based classification for changing scenes. The changing scenes bring a series of new problems to the classifer which trained in the old scene: (1) The distribution of sample set in the old and new scene is different, this lead to the classifier which trained in the old scene can not adapt to changing scenes, so the performance is not well; (2) Although there exists some correlation in the sample set of old and new scenes, the classifier which trained in the old scene can not completely reflect the intrinsic characteristic of the new scene, it leads to the detection performance in the new scene decrease. These difficult issues make pedestrian detection method based classification to be recognized as challenging technical problems.This paper argues that the similarities and differences between the changing scenes can handle in feature selection and classifier design two levels. So this paper studies pedestrian detection method based classification from the feature extraction and classifier design two sides, and provides technical support for pedestrian classification detection system. The main work and innovations are as follows:1)Considering the problem that the classifier which was trained in the old scene can not adapt to the changing scenes, propose a classification optimization model method for changing scenesThe detection performance of the classifier which was trained in the old scene does not work well in the new scene. If we rebuild the classifier in the new scenes every time, it is difficult to obtain training samples in a short time, and the training time is too long that it is difficult to meet real-time detection application. This paper makes the best of the existing classification structure and adjusts the decision variables of classification model to adapt to the new scenes for pedestrian detection. First we train a cascade classifier in the old scenes and maintain its cascade classifier structure meanwhile, and then convert it to a threshold decision vector optimization problem, and then use a small number of representative samples in the new scenes to adjust the key parameters to adapt to the new scene dynamic. In addition, in order to further speed up the detection rate in the new scenes, a general ternary detection pattern based cascade classifier structure was proposed. Experimental results verify the proposed classification model optimization method can achieve good detection performance and fast detection speed.2)Considering that the sample set of old and new scenes have the same feature space, a feature transfer model in changing scenes was proposed.The sample set of old and new scenes are generally in the same feature space. There must exist certain common between the two feature set. If we can take advantage of the common, it can reduce the difficulty of feature extraction in the new scenes. This paper proposes a feature transfer model in the changing scenes which transfer the high-level feature in the old scenes to the new scenes for pedestrian detection. The feature transfer model extracts the high-level feature in the old scenes using sparse coding algorithm at first, and then let it be as the input for feature extraction in the new scenes to exact the sparse feature for new scene samples. Experimental results show that the feature transfer model can achieve better detection performance in the changing scenes.3)Considering that there exist some similar samples between the sample set of old scenes and new scenes, a classification transfer model in changing scenes was proposed.There exist some similar samples between the sample set of old scenes and new scenes. Due to there are only small amount of samples in the new scene, and it can not to train a reliable classification model, so we need to make best of these similar samples in the old scenes. This paper proposes a pedestrian classification transfer model in changing scenes with feature extraction and classifier design. The model consists of two parts: sample screening method based manifold learning and classification model design based transfer learning. First, the method describes the samples in the old scene and the new scene visually, and then selects several samples in the training set, which are very similar to the samples in the new scenes, then he selected samples are merged into the test set to expand the size of training set, last rebuild a new classification model by the new training set and large amount of auxiliary data. In addition, the classifier designs with the idea of transfer learning and adjusts the sample weight through special strategies. Experimental results demonstrate the effectiveness of the proposed pedestrian classification transfer model.