The Research Of Pedestrian Detection And Recognition Under Complex Background

With the increase of intelligent demands in vehicle assisted driving,video monitoring system,intelligent robot,the pedestrian detection technology have become one of the core applications in those fields.Under the complex environment,there are some problems in using a single feature to describe pedestrians,such as low detection and recognition rate and low real-time performance.Therefore,this thesis focuses on these issues,the main research contents are as follows.Under the complex environment background,the extracted pedestrian features have more interference information,which leads to the low recognition rate of pedestrian detection.To solve this problem,this thesis proposes a method based on HOG(Histograms of Oriented Gradients)feature describing pedestrian gradient and ULBP(Uniform Local Binary Pattern)feature series fusion of texture information,which effectively improves the recognition rate of pedestrian detection.However,the vector dimension with multiple features is too high,which leads to the increase of pedestrian detection time and slowdown detection speed.Through the analysis of the experimental process,this thesis proposes a method,which utilizes the principal component analysis algorithm to reduce the dimensionality of HOG feature in high latitude to form HOG-PCA+ULBP feature.The results of the study show that the improved algorithm can effectively improve the pedestrian recognition rate and detection speed.It is found in this thesis that HOG feature and ULBP feature produce different recognition effects when they recognize pedestrians and simple multi-feature tandem fusion method cannot give full performance to pedestrian detection.Due to the problems and shortcomings of multi-feature serial fusion,this thesis proposes a pedestrian detection algorithm that combines the weak classifier with different weight coefficients to form a strong classifier.By training HOG weak classifier and ULBP weak classifier on test samples,a pedestrian detection algorithm based on the combination of weak classifier and different weight coefficients is proposed.In the first stage of pedestrian detection with this classifier model,it is found that the ULBP weak classifier can effectively eliminate the non-pedestrian target window,the strong classifier will separate and identify the remaining windows to be detected containing pedestrian targets.This method can effectively save detection time,improve detection speed and pedestrian recognition rate.When the pedestrian detection window is fused,the final output pedestrian detection window may not eliminate the smaller false detection windows.In order to solve this problem,this thesis proposes a method to improve the traditional greedy non-maximum suppression algorithm.Firstly,the traditional NMS window fusion algorithm is utilized to screen out the pedestrian target window,and then the size of the maximum confidence pedestrian detection window and the confidence threshold in each group is compared.Finally,the smaller pedestrian detection windows detected by mistake are removed according to the size relationship.The results of the experiment by Matlab show that the improved NMS algorithm eliminates the smaller detection window that is misdetected successfully.As a conclusion,the improved algorithm proposed in this thesis effectively improves the pedestrian detection recognition rate and detection speed under the complex environmental background.The verification of the experiment proved by Matlab shows that the weighted weak classifier pedestrian detection algorithm improves the accuracy by 20%,and the detection time of the single image is shortened by about 0.04 s.Also,the multi-feature fusion algorithm is better than the single-feature description algorithm in pedestrian performance.