Research On Multi Task Joint Training Method Of Human Perception

In commercial,security,military,transportation,medical and other practical application scenarios,the human body is always the most closely concerned object.There are many research directions taking the human body as the object,including human body detection,human body instance segmentation,human body key point detection,human body part segmentation,human body part detection,human body dense point prediction and other technical directions for all dimensions of the human body.The predicted results of these studies contain independent onedimensional information at all levels of the human body.At present,each sub direction of the human body is developing well vertically,but in practical application,it is also necessary to integrate multiple task results,hoping to obtain more comprehensive and multi-dimensional human perception results.Because the multi task model will lead to the hardware pressure of training and the low training efficiency,the multi task joint training technology has become a better solution.However,there are also some deficiencies in the field of joint training,such as few tasks,task competition,limited accuracy and so on.Based on the relationship between multiple tasks of human body,this paper proposes an end-to-end solution that can realize the parallel prediction of multiple tasks.At the same time,on this basis,in-depth optimization is carried out for the problems of precision loss,lack of supervision and single task in multi task joint training,so as to finally realize a more comprehensive and better performance human perception joint training method.Firstly,the data sets of multiple tasks are filtered and fused to construct a joint data set of human object multidimensional annotation.Then,the network design and experiment of perceptual one-dimensional task are carried out,and the initial benchmark accuracy system is established,which is used as the control group of multi task joint training method and multi task single training method.Then,this paper selects the technical scheme route of end-to-end joint training for multi task training,obtains the accuracy benchmark of the basic multi task network,introduces the whole network pre weight information supervision and segmentation information supervision respectively for the lack of supervision in the scheme,puts forward the weight allocation strategy for task competition,and increases the human body part detection task for less task dimension,The attention mechanism and spatial pyramid are introduced to enhance the head network.Finally,this topic integrates all improvements and optimizes the human multi-dimensional perception network,which significantly improves the overall accuracy.Among them,APb is improved to 53.1%,APm is improved to 49.6%,APp is improved to 52.9%,APkp is improved to 65.2%,and APd is improved to 60.9%.At the same time,newly added APh is also reached 34.5%.