Pedestrians Detection Algorithm Based On Multi-Features Fusion

Nowadays, the pedestrians detection is one of the hotspots in the field of computer application. It plays a very important role in the fields of intelligent monitoring systems, human-computer interaction, intelligent Vehicle assisted navigation technology, computer vision and pattern recognition, which has very high scientifical and economical value. So, researching on the pedestrians detection is very significant.This paper mainly studies the method of detecting pedestrians in still images. Firstly, it uses HOG and HOG_LBP（a mixed feature） to express the charactristics of pedestrians. Then, it selects the feature vectors of the extracted HOG features with Rough Set theory to generate reducted feature-vector. Lastly, training the Support Vector Machine（SVM） with the extracted feature vectors to detect pedestrian predict the pedestrian’s test data set.The main content of the thesis is:（1） Using the images in the INRIA pedestrian library to establish two sample sets. In the SVM classifier training, to study the impact of the size of sample set to SVM classifier results.（2） The HOG, 8,1LBP and u 28,1LBP feature were extracted respectively from the images of concentrated samples to get two kinds of hybrid feature matrix by feature fusion. With HOG 8,1LBP and HOG_u 28,1LBP, the ability of characterizing of the pedestrian features was demonstrated with different feature matrixes.（3） The HOG feature matrix was constructed into decision table with discrete values to study the method of attribute reduction of feature vectors using rough set theory, and the reduced feature vectors were coupled with 8,1LBP and u 28,1LBP feature to get mixed features: HOGR₈,1LBP and HOGR_u 28,1LBP.The experiment indicates: under the same conditions, the bigger the size of the sample set, the higher the final accuracy of the classifier. In the larger sample set, the detection result using mixed features was higher than that using single feature. Using the method based on the Rough Set theory to complete attribute reduction of feature vectors, the detection result was not affected, however the dimension of feature vectors was reduced and the detection efficiency was improved.