Continuous Target Recognition Based On Deep Learning

Human beings can easily recognize moving or deformed objects,but there are many problems in computer recognition: traditional recognition algorithms are affected by background light changes and shape changes when detecting continuous object;traditional recognition algorithms perform well in static target recognition,while the accuracy of continuous target recognition is not high enough,and the real-time performance can not be guaranteed.To address those issues,this thesis uses deep learning algorithm to recognize continuous object from different backgrounds,and compares this algorithm with other classical algorithms to test and verify on different data sets.The main contents and achievements of this thesis include:Firstly,Based on the motion estimation algorithm of continuous video frames,the regions of interest in video frames are detected.Taking CORe50 data set as the research object,three benchmarks of dynamic handheld object recognition are proposed,which are consolidated on the basis of known classes,unknown classes and known objects respectively.They are used in subsequent dynamic handheld object recognition.Then CNN features and SIFT features are fused to obtain comprehensive features.With AlexNet network and VGG network training,the classifier is trained to get the recognition results.Experiments show that the recognition accuracy of the proposed algorithm is higher than that of traditional recognition algorithms.Secondly,In order to ensure the real-time performance of continuous object recognition,an SSU-SGD optimization algorithm with different step sizes is proposed.On this basis,three benchmarks for dynamic handheld object recognition are updated,and then AlexNet and VGG networks are used for training and recognition.Experiments show that the recognition accuracy is improved significantly,up to 70.14%.In addition,the recognition speed based on SSU-SGD algorithm is also improved significantly,which basically meets the real-time requirements of dynamic object recognition.Finally,Based on the idea of discriminative learning and unsupervised domain adaptation,a discriminator domain adaptive network(DDAN)is proposed,which includes appearance adaptive network and feature adaptive network.Domain discriminator is added to the shared feature layer of feature adaptive network.Domain discriminator is deceived by continuously learning the feature output of source domain and target domain.SSU-SGD algorithm is used to optimize the network,and domain discriminator can converge quickly and achieve the result of dynamic object recognition.Experiments show that the recognition accuracy of DDAN in HMDB-A dataset is up to 73.05%,which basically meets the requirements of continuous object recognition.The performance of IoU is also improved,and compared with other data sets,DDAN network can also achieve certain recognition results on other data sets.In summary,the main work of this thesis is to use deep learning algorithm to achieve continuous object recognition,and validate it on different data sets,and get better recognition accuracy,and the training speed basically meets the requirements of real-time.