Multi-scale aware pedestrian detection method based on improved full convolutional network

Abstract

The present invention relates to the field of pedestrian detection, and particularly relates to a multi-scale aware pedestrian detection method based on an improved full convolutional network. Firstly, a deformable convolution layer is introduced in a full convolutional network structure to expand a receptive field of a feature map. Secondly, a cascade-region proposal network is used to extract multi-scale pedestrian proposals, discriminant strategy is introduced, and a multi-scale discriminant layer is defined to distinguish pedestrian proposals category. Finally, a multi-scale aware network is constructed, a soft non-maximum suppression algorithm is used to fuse the output of classification score and regression offsets by each sensing network to generate final pedestrian detection regions. Experiments show that there is low detection error on the datasets Caltech and ETH, and the proposed algorithm is better than the current detection algorithms in terms of detection accuracy and works particularly well with far-scale pedestrians.