Research On Pedestrian Detection Algorithm Based On Image Multi-feature Extraction

In recent years,pedestrian detection technology has been widely used in the field of artificial intelligence.Because pedestrian target has changeable posture and is affected by environment easily,it is different from other rigid targets.How to efficiently and accurately detect pedestrians from complex scene images has become a hotspot and difficult point in current research.In order to improve the effectiveness of pedestrian detection in complex scenes.pedestrian detection technology was studied by this thesis based on image multi-feature extraction algorithm.The main contents of this research are as follows:(1)In view of the difficulties faced by pedestrian detection in practical application,the representation ability of the currently widely used feature extraction algorithms on pedestrians was systematically analyzed by this thesis.The advantages and disadvantages of the mainstream classification algorithms on pedestrian classification and recognition was studied,which provides a theoretical basis for the subsequent research on pedestrian detection algorithms.(2)Aiming at the problems of traditional HOG(Histogram of Oriented Gradient)feature extraction speed is slow and data dimension is large,rapid HOG feature extraction was studied.The HOG integral graph was calculated by using the principle of integral graph,which improves the speed of HOG feature calculation.The feature dimensionality reduction algorithm was used to process the HOG feature matrix dimensionality reduction,eliminating the redundant data and reducing the calculation amount.Experimental results show that the adoption of fast HOG feature has a higher detection efficiency for pedestrian detection.(3)Aiming at the problem that a single feature could insufficiently represent pedestrians in some complex scenes,a pedestrian detection algorithm combining fast HOG feature and Gabor feature was studied.The Gabor features of the samples were extracted by Gabor filter.A hybrid feature classifier was generated from fast HOG feature and Gabor feature based on serial fusion method.The experimental results show that,compared with the classifier composed of single features,the classifier composed of mixed feature can effectively reduce false positive rate and false negative rate in some complex scenes,and improve the correct rate of pedestrian detection.(4)A multi-feature weighted cascade classifier was proposed to solve the problem that the features with different representation ability have the same weight in serial fusion.Adaboost classification algorithm was used to respectively train the Gabor weak classifier and the fast HOG weak classifier.The weak classifier was weighted to build a strong classifier,and the multi-feature weighted cascading classifier was obtained through training.Finally,two detection window fusion algorithms are used to analyze the overlapping windows generated by the classifier.The results show that the non-maximum suppression algorithm can effectively screen the overlapping windows,and it is more accurate to mark the pedestrian position compared with the window area comparison method,which can effectively reduce false negatives rate of pedestrians.By building an experimental platform and designing an experimental scheme,the performance of the pedestrian detection algorithm studied in this thesis was systematically verified by experiments.The results show that the multi-feature cascade classifier composed of fast HOG feature and Gabor feature has higher detection accuracy and faster detection speed in pedestrian detection,which can effectively reduce false positive rate and false negative rate in pedestrian detection on complex occasions.