Research On Pedestrian Detection Algorithm For Driverless Vehicle Vision System

With advances in technology and productivity,driverless cars have become no longer strange.But it also needs more comprehensive technical support to enter people's lives.Pedestrian detection has always been a difficult part of driverless cars.How to effectively avoid pedestrians is the problem that researchers most want to solve.Pedestrian attitude changes,different pedestrian dress,complex background images,changing lighting conditions and real-time testing are the current difficulties in pedestrian detection.The purpose of this paper is to propose a system of pedestrian detection algorithms for driverless cars in view of the above difficulties.At first,for the original HOG extraction method can not solve the problem of aliasing in adjacent areas,an interpolation method is proposed,and the three-line interpolation is applied to the cumulative histogram correction.Comparing the support vector machines in three different situations,the SVM training method is used to find the optimal support vector machine,and a method to generate strong classifiers is proposed.Secondly,the pedestrian database is the sample basis of pedestrian detection.The sample database of this trial is made by using two databases of MIT and INRIA respectively.LIBSVM software package is used to carry out classifier experiments.During the experiment,the most suitable kernel function is found through the comparison test,and then the best sample INRIA sample is obtained by using the kernel function to find the two samples.Finally,a method to evaluate the performance of support vector machines by the ROC curve is proposed.Then,based on the principal component analysis,this paper puts forward a fast feature extraction method that combines PCA and HOG features.According to this method,experiments were conducted to find the most suitable p value for the experiment in the paper,and the optimal p value was used to compare HOG characteristics with the time and recognition rate of HOG-PCA pedestrian detection.The feasibility and speed of this method was verified.Finally,introducing the LBP feature,a multi-level combination method of HOG features and LBP features is advanced,and the computational recognition frame occupancy ratio is added in the secondary step to further filter out the pedestrian target in this experiment.Designing the sensor system of the driverless car,building an experimental hardware platform for the driverless car vision system,writing a software interface to drive the test.And perform pedestrian image and complete the comparison experiments by using this platform.Verify the applicability of this system to pedestrian detection in driver-aware systems.