Pedestrian Detection Method Based On Multiple Feature Decision Level Fusion

Pedestrian detection is a hot topic in the field of computer vision in recent years.Pedestrian detection technology is one of the key technologies of the auxiliary driving system.It can detect the pedestrians in the surrounding environment in real time,and then prompt the driver to react or to make emergency brake in a dangerous situation to reduce or avoid traffic accidents.Pedestrian detection is the detection of road or public area in natural scene,which is easily affected by illumination,weather and motion blur.Therefore,pedestrian detection is a very worthy and challenging topic.A pedestrian detection method based on SVM-DS multi-feature decision fusion is designed in this paper.The conclusion shows that this method has some advanced and potential application value.First of all,two features of HOG and LBP are extracted,and then the posteriori probability output of each feature which is used as the basic reliability assignment is obtained by using probabilistic SVM.After that,according to the DS evidence theory,the data fusion is carried out based on the fusion rules of matrix analysis,and the decision rules are made to realize the output of pedestrian detection results.The experimental results show that this method can effectively realize the fusion of pedestrian HOG features and LBP features at the decision-level,the comprehensive advantages of HOG feature to depict pedestrian profile and of LBP feature to depict pedestrian exterior texture are fully utilized,and improved pedestrian detection performance is obtained.The primary contents of this thesis are as follows:1.Two kinds of feature extraction methods,HOG and LBP,are studied.Through the experiment of single feature detection and comparison,the feature extraction scheme is determined.2.A Pedestrian Detection method based on feature amplification and SVM is designed and achieved.Complementing the limitations of the HOG feature descriptor using the LBP feature,the experimental results of INRIA data sets show that the detection rate is 94.59%.But Pedestrian Detection method based on feature amplification and SVM does not eliminate theconflict or prejudice caused by single features on the same recognition object.3.A pedestrian Detection method based on probabilistic SVM and DS evidence theory decision-level fusion algorithm is designed and achieved.DS evidence theory can eliminate the conflict between single feature and the same object detection result.The detection rate of this method in INRIA data set is 95.86%,which can effectively improve the accuracy of pedestrian detection.