Research Of Pedestrian Detection Methods Based On Features Fusion And Multi-kernel Learning

Traffic safety issue have became a hot social topic. In order to solve this problem. All sectors of society are committed to the development of a variety of driver assistance systems to improve driving safety. Pedestrian detection is an active safety technology, through the analysis of the driving environment, to early warning of the possibility of traffic accidents, active intervention on car (emergency brakes), it can effectively reduce the possibility of traffic accidents.This paper studies the common pedestrian features and classification algorithms. We found that the performance of single feature combined with the single classifier is limited.To come up with high real-time performance and strong robust pedestrian detection algorithm. This paper adopts multi feature fusion and the fusion of multiple classifiers.As the low rate of detection of the method only using single-feature, To make use of multi-source information feature, we fusion some low-level features (color、gradient、 histogram) and multi-level oriented edge energy feature based on the linear discriminant analysis of linear weighted fusion strategy. Features used for pedestrian detection are high in dimension. We use principal component analysis to reduce the dimension of features, and make the detection algorithm runs faster. It overcome the influence of the high dimensional features which reduce the real-time of pedestrian detection. Features can be calculated fast by integral image technique. The robustness and real-time performance of pedestrian detection system have been strengthened. Histogram intersection kernel support vector machine have the advantage of fast classification and high accuracy in object recognition. It can be used for further enhancing the system real-time performance. The experiments show that the proposed algorithm has faster detection speed and higher precision than the classical algorithm HOG+SVM.Single kernel SVM classification performance is limited. We adopts the method of multi-kernel learning for pedestrian detection, We synthesis a new kernel function by the linear combination of HIKSVM, Gauss SVM and polynomial kernel SVM.It can be used to classify. Experiments prove that multiple kernel learning algorithm has enhanced detection rate. Due to high computational complexity, The real-time performance of pedestrian detection system is poor. Finally, we summarize the research work of the full text, points out the problems and difficulties faced by the pedestrian detection, prospects the future works.