The Pedestrian Detection Based On Spatio-Temporal Interest Point And Histograms Of Oriented Gradients

In recent years, the intelligent monitoring system on the embedded devices, due toinvolving in people’s safety and vital interests, increasingly arouse people’s attention. In thesystem, pedestrian detection is the base of the pedestrian tracking and behavior recognition,and its importance is evident. However, in the application of computer vision, the pedestriandetection is still an impotant topic, and its research methods include image processing,artificial intelligence, and so on. Therefore, the research of the pedestrain detection has valuesboth in the applicaion and in the theory.Pedestrian detection generally includes two steps: pedestrian object extraction andclassification. The pedestrain object extraction relates to pedestrian object segmentaion andregion extraction. As a method for pedestrian object segmentation, Kim codebook methodoften has long modeling time in practical application, so demands of real-time applications isalways not meted. For the simple implement, the frame difference method is often used toextract pedestrian object region, but when the environment has large changes, the mehod cannot stably express scene. The histogram of gradient detection in the whole image is a classicalmethod, but detection in the whole image leads to long detection time and low accuracy insmaller noise scene.Aiming at the disadvantages of the above methods, this paper has done the followingwork:1) According to the problem that the Kim codebook method has the long modeling time,base on the frame difference codebook method, this paper changes the learning threshold toget the realtime update in the local area. The experimental results show that the local upadteKim codebook has the lower background modeling time and higher accuracy than the basiccodebook method in a certain environment.2) According to the problem that the frame difference method has low stability in somescenarios, this paper proposes a strategy using spatio-temporal interest points instead of thewhole-image. By adding optical flow method to compare the extraction based on framedifference with the extraction based on spatio-temporal interest, the experimental results showthat spatio-temporal interest points can replace the scene steadly in a certain environment.3) Aiming at the problem that the histogram of gradient detection in the whole image hasthe long detection time in the low noise scene, this paper presents a strategy using subimageinstead of whole-image, and proposes the pedestrian detection algorithm based onspatio-temporal interest point and the histogram gradient. The experimental results show that the subimage detection can improve the accuracy, and reduce the detection time in low noiseenvironment.