Local Feature Learning For Pedestrian Detection Of Static Image

Pedestrian detection technology is promising and has broad applications such as video surveillance, intelligent robot, driver intelligent assistance, pedestrian traffic analysis and so on. However, detecting humans is one of the most challenging tasks in machine vision owing to their variable appearance and the wide range of poses, a change in terms of Angle, different illumination and cluttered backgrounds. Meanwhile, pedestrian detection systems are required to have high precision and the real-time performance.Because a good method of pedestrian detection of static image is the prerequisite of the pedestrian detection system, On the basis of sufficient investigation on the related algorithms of pedestrian detection, this paper studies the model of classical algorithms of pedestrian detection. This paper proposes three new methods of pedestrian detection, which achieve good performance.The main contributions of this paper are as follows:1. This paper makes a brief introduction and summary of the current situation and the exiting methods in pedestrian detection field. Moreover, this paper discusses the current problems, difficulties and development which exist in the exiting methods of pedestrian detection.2. This paper introduces the theoretical basis and key technologies, which are based by the new methods of pedestrian detection in this paper.3. A pedestrian detection method based on Multi-scale Block Local Binary Patterns (MBJLBP) features and Histogram Intersection Kernel SVM (HIKSVM) is proposed. Results of the comparative experiments show the superiority of ours to some classical algorithms such as HOG+LinearSVM.4. A pedestrian detection method based on Subset-Haar-like template features is proposed. The target features are an intermediate layer features generated by using Subset-Haar-like template filtering low-level Aggregated Channel Features (ACF). Our detector is modelled by the target features in combination with a boosted decision forest Our method can filter out non-person window more effectively and has an even lower false positive rate per image.5. A pedestrian detection method based on weighted features of Subset-Haar-like template features is proposed The target features are constituted by the weighted sum of original Subset-Haar-like features, while LDA are used to learn weighting coefficients. The results based on INTIA dataset indicate the effectiveness of our method.