Research Of Pedestrian Detection In Image

Pedestrian detection is to determine whether pedestrians are present and locate their location in the input image or video sequence.Pedestrian detection is widely used in the fields of intelligent vehicles,automatic monitoring,human-computer interaction,virtual reality and other fields.At the same time,pedestrian detection requires high precision and real-time speed,but pedestrians has both property of rigid and flexible objects,its appearance is easy to be affected by wear,scale,occlusion,gesture and perspective,making pedestrian detection become one of the hot and difficult research direction in the field of computer vision.This paper focuses on studying pedestrian detection algorithm in the image.The main research works are listed as follows:(1)In order to overcome the problem that integral channel features has redundant information and slow detection speed in multiscale pedestrian detection,an improved integral channel features for fast multiscale pedestrian detection is proposed.Firstly,fast feature pyramids is used to compute multiscale channel features,which avoid computing the same features at multiple locations and scales.Then the detection windows are divided into cells and blocks to make the overall description of the image and reduce the redundancy,replacing the original method of random location and size.Finally,obtaining the sums of pixels in cells and blocks as the pedestrian features and then the features are classified by soft cascade Adaboost.Simulation experiment results show that the proposed method gives improved detection rate and 11.3 times speedup compared with the original algorithm and runs at13.54 fps on 640x480 images.(2)ACF algorithm has more false detection windows,so this paper proposes a coarse-to-fine cascaded pedestrian detection algorithm.Firstly,ACF is employed as the coarse detector,then improved channel features is extracted from these candidate windows to filter out false detection windows.When improved channel features,the algorithm learn PCA filter banks from each channel,instead of learn PCA filter banks from the image and convolution maps,then the filter bank convolution with channels to improve feature discrimination capability,instead of the original two layers convolution to reduce feature dimension,finally a pooling operation is applied to convolution maps to reduce the feature dimension.Simulation experiment results show that comparison with the original ACF algorithm,the proposed method have less false detection windows and detection rate on INRIA and Caltech database increase by 3.8% and 10.3% respectively.(3)In order to overcome the problem that in the ACF algorithm the pedestrian appearance is unstable,which resulting in a decrease in detection rate.Pedestrian detection based on fast edge detection and real adaboost is proposed.Firstly,features in the image patches isextracted,then feed into the trained random forest classifier to identify whether the pixels are edge pixels in patches to obtain a more stable edge map.To obtain a more stable pedestrian contour information,then use the edge map as pedestrian contour channel to replace the normalized gradient magnitude in the ACF algorithm.Finally,the real adaboost classifier with higher accuracy is used to improve the classification accuracy.Simulation experiment results show that comparison with the original ACF algorithm,the proposed method have less false detection windows and detection rate on INRIA and Caltech database increase by 5.1% and14.8% respectively.