Research And Application Of Joint Network For Person Detection And Attribute Analysis

Regarding recognition of human attributes in the applications,it is necessary to first identify and detect the human and then analyze their attributes for each detected object.At present,the mainstream approach is to train two independent deep convolutional networks to finish human detection and attribute recognition tasks separately.The disadvantage of the previous approach is that these two networks are trained independently,which would take away dual training time and computing power,they can't share the internal information learned from different tasks.Besides,different appearance attributes of humans are treated as independent,which would lead to the inefficient use of prior knowledge.For example,one that wears a skirt is generally female,and formal clothes often conflict with shorts.This empirical knowledge will help improve the performance of attributes recognition.Finally,weighted summation for losses of different attributes,which is the usual method to deal with multi-object optimization problems,can't avoid competing conflicts between multiple tasks.In response to the above issues,the following research is carried out:First,a joint learning network framework Fine Gained Multi-Attribute R-CNN(FGMA R-CNN)for end-to-end object detection and multi-attribute recognition is proposed to realize the detection of multiple objects and the simultaneous reasoning of multiple attributes of objects.This method is based on Faster R-CNN to complete the analysis of human attributes by integrating a branch that predicts human attributes.The two tasks of human detection and multi-attribute analysis share the weight and characteristics of the backbone network so that to complete human detection and attribute recognition at the same time just use one network.Compared with the step-by-step model of object detection and attribute analysis,not only ours can obtain good generalization ability,but faster.Secondly,the correlation between human attributes is analyzed by calculating the confidence and lift between attributes.This attribute correlation is used as prior knowledge to guide the network to learn the correlation of human attributes to avoid the network produces results that violate prior knowledge and improve the performance of individual attribute analysis caused by imbalanced training data.Then,a multi-objective optimization algorithm based on multi-gradient descent is proposed,the algorithm is suitable for a two-level multi-task network of object detection and attributes analysis.By looking for the Pareto optimal solution of the network,the optimization problem in the network training process is transformed into a multi-objective optimization problem.This method can effectively avoid the conflict between tasks,and further improve the generalization performance of the network.Finally,we apply FGMA R-CNN to the elevator scene to verify its engineering feasibility.We develop an intelligent visual recognition system that can exact attribute information for persons.And the adaptability in special application scenarios of the proposed method is studied with regard to indicators including network scale,computing efficiency,accuracy,and hardware requirements,and so on.