Person Re-identification Based On Dual Attention And Part Drop Network

Pedestrian re-identification,also called as pedestrian retrieval,is a technology that uses machine learning algorithms to retrieve a given pedestrian from a large amount of video surveillance data.Input several pedestrian images into the pedestrian re-identification system,and search for the same pedestrian captured by different cameras through a well-designed algorithm.Combined with computer vision technologies such as pedestrian tracking and pedestrian detection,pedestrian re-recognition can well solve the limitations of the large-scale camera network and can be used in a wide range of fields such as intelligent video analysis,intelligent public security,and intelligent tracing.Due to the certain differences between monitoring equipments,and pedestrians are not a fixed rigid body,their appearance is easily affected by clothing,scale,occlusion,posture changes and viewing angle variations,making pedestrian re-identification a very challenging subject in research field.In order to reduce the impact of occlusion on person re-identification,this paper proposes an algorithm based on Dual Attention and Part Drop Network(DAPD-Net)to extract the discriminative features of pedestrians.The dual attention module(Dual Attention Module)in the network enables the basic neural network Res Net-50 to focus more on the foreground pedestrian target and ignore the background interference;the Part Drop Branch divides the feature map into multiple parts.Then randomly occlude one of the them and learn the remaining components to obtain strong robust features against occlusion;Middle Layer Branch is used to extract discriminative middle-level semantic information features;Global branch which is used to encode global high-level semantic information features.By reasonable combination of each module branch in the network,a more discriminative feature representation can be learned.Using Tri Hard loss and Label-smoothed cross-entropy loss to change the metric learning of human re-identification.The method in this paper has conducted a large number of experiments on multiple public benchmark datasets,and the results fully show that the method in this paper is better than many state-of-the-art methods and the performance has also been significantly improved,thus proving the advanced performance and effectiveness of the algorithm proposed in this paper.