Research On Human Motion Micro-doppler Image Quality Assessment

In order to cope with the trend of population aging,it is necessary to employ Internet technology to monitor the activities of elderly people living alone,so as to prevent injuries or other accidents.As an important tool in indoor monitoring at present,radar can be widely used in the field of micro-moving target recognition since it is not affected by illumination and has a strong ability to capture subtle movements.However,the image quality of radar microDoppler image is always degraded due to the environmental noise or other signals,thus the human motion monitoring and recognition effect is reduced.Therefore,in order to remove the interference of lower quality micro-Doppler images,the research on human motion microDoppler image quality assessment is deployed.Human motion micro-Doppler image quality assessment algorithm is necessarily studied based on a human motion micro-Doppler image quality assessment database,nevertheless,there is no such public database available.Hence,a Human Motion Micro-Doppler Image Quality Assessment(HMMDIQA)database is constructed in this paper.The main research contents are as follows:(1)A HMMDIQA database is constructed.Firstly,1230 micro-Doppler images of human motions including boxing,walking,running,crawling and stooping are captured and generated.Then,white Gaussian noise at five levels and threshold noise at three levels are added to generate corresponding distorted images.The constructed HMMDIQA database is consist of1230 reference images,9840 distorted images with subjective scores.The database is proved to be reliable and challenging through experiments.(2)A HMMDIQA algorithm with principal component analysis-based subspace feature enhancement is proposed.This algorithm is a multi-task learning algorithm,wherein,the quality prediction task is the main task,while the subspace feature prediction task is the auxiliary task to enhance the features of human motion micro-Doppler images.In this algorithm,the two tasks share a same feature extractor.Meanwhile,in order to improve the characterization ability of the feature extractor for micro-Doppler image features,an asymmetric residual module is added to the feature extractor.The algorithm can learn the mapping relationship between human motion micro-Doppler images and quality scores more completely by realizing the subspace feature and quality score prediction tasks of microDoppler images,thus can increase the accuracy of the quality prediction task.The HMMDIQA database constructed in this paper has subjective quality score,which is a valuable work in the field of human motion micro-Doppler image research.The proposed HMMDIQA algorithm with principal component analysis-based subspace feature enhancement has a SROCC value of 97.34% and a PLCC value of 97.79% on the HMMDIQA database,which is 2.7% and 2.77% higher than the baseline network,respectively.