Method Research On Pedestrian Detection Based On Visual Attention Mechanism

Pedestrian detection is an important research topic in computer vision,and plays an important role in vehicle assisted driving,robot navigation,and intelligent monitoring.In recent years,with the understanding of people's vision,visual attention has made a major breakthrough in the application of object detection.Therefore,applying visual attention mechanism to pedestrian detection has important value and significance.This article mainly completed the following work:1)Primary visual features and semantic features.Firstly,we improve the GBVS visual attention model to enhance the saliency of pedestrians according to the external characteristics of pedestrians,the visual attention model based on graph theory is improved to further enhance the pertinence of pedestrians in the scene.Secondly,the model of regional skin color is used to calculate the saliency of skin color semantic features.Finally,we use machine learning to extract head-shoulders features,and use Haar-like features to identify the head-shoulders and as SVM to train the classifier.2)Saliency map fusion.We use the Laplacian pyramid to fuse the salient maps to create a static visual attention model.Firstly,establish a Gaussian pyramid for the primary visual saliency map,the skin color saliency map,and the head-shoulders saliency map.Secondly,the region-based fusion rules are used to fuse the saliency maps at the same scale and obtain gradient pyramids.Finally,total saliency map is obtained using image reconstruction.3)Focus of attention selection and shift.We process the saliency map with focus of attention(FOA)selection and shifting with a DPM to complete the pedestrian detection.Above all,we calculate the largest pixel value on the saliency map,and we use the method of region-growing method to obtain the FOA.Furthermore,we calculate the HOG features of the FOA and then send them to the LSVM for determination to complete the FOA shift.Experimental results show that our static visual attention model performs better than other models,the accuracy of pedestrian detection reach 92.78% by using Laplacian fusion strategy on the INRIA Dataset.4)We present a method to improve pedestrian detection using a visual attention mechanism with deep learning.Firstly,a top-down visual attention model is created through short connections and a visual saliency map is generated.We multiply the visual saliency map with the input image to generate multiplied visual salient image.Finally,the multiplied visual salient image is sent to the detection network to detect pedestrians.Experimental results demonstrate that the proposed achieves state-of-the-art performance on Penn-Fudan Dataset with 91% detection accuracy and it achieves average miss-rate of 15% on the INRIA Dataset and achieves average miss-rate of 28% on the Daimler Dataset.