Research On Object Detection And Tracking Based On Deep Learning

The object detection and tracking has an extensive application in military and civilian fields,and has become the focus and one of the most essential directions in computer vision.The emerging of deep learning has greatly promoted the development of computer vision.To improve the performance and expand the applications in object detection and tracking,this paper has carried out the object detection and tracking research based on deep learning,which mainly includes the following four achievements:(1)Due to the deficiency of deep neural networks based on regression in the small object detection,the multi-view object detection method based on deep learning was proposed.The working principle,the model structure and the model implementation were elaborated and the experiments were carried out on the 106 small object datasets selected from VOC2007 datasets.Through qualitative analysis and quantitative analysis,the performance improvement of YOLO(You Only Look Once: Unified,Real-Time Object Detection),YOLOv2(based on an updated version of YOLO)and SSD(Single Shot Multibox Detector)with the multi-view models were obviously improved.In the experiment,the mAP of multi-view YOLO has been increased by 14.3% compared with that of YOLO,the mAP of multi-view YOLOv2 has been increased by 7.4% compared with that of YOLOv2 and the mAP of multi-view SSD has been up by 13.1% compared with that of SSD.Meanwhile,the multi-view model can be universally applied in the deep learning object detection methods based on regression and offers a well solution for the deep learning object detection methods based on regression to improve the performance in small object detection.(2)In order to transfer the advantage of object detection based on deep learning to object tracking,an object tracking method based on deep learning was proposed.The method made the best of the advantages of deep learning in feature expression by adopting SSD based on regression to extract the candidate objects and screen the objects by combing the color histogram feature and HOG feature to implement object tracking.To improve the performance of object detection model,the paper constructs a multiview object search map so that different scales of object detection can be achieved on just one map.Meanwhile,the paper has selected eight typical tracking video sequences on the standard object tracking benchmark and six typical tracking algorithms in conducting comparative experiments.The results show that the above-mentioned algorithms have generally better tracking effectiveness than other algorithms and have better robustness for such influence factors as object change,size change,rotation,illumination change and complicated background clutters.(3)In order to utilize deep learning method on object detection and tracking for infrared and visible spectra,a decision-level fusion detection method and a decisionlevel fusion tracking method based on deep learning were proposed.First,a parameter transfer model for deep learning model was proposed.And then a pre-training model for infrared object detection was extracted from a visible object detection model based on deep learning and was fine-tuned on a collected infrared image dataset to obtain an infrared object detection model based on deep learning.On this basis,a decision-level fusion model for infrared and visible detection and a decision-level fusion model for infrared and visible tracking were estabilished and the corresponding experiments were implemented,respectively.The experiment results show that the proposed methods of decision-level fusion for infrared and visible spectra are are superior to the single band methods on object detection and tracking in terms of their performance and their robustness.Meanwhile,the proposed methods offer an excellent solution for multi-band fusion detection and tracking based on deep learning.(4)An image grapping platform of the infrared and visible images was established and a making scheme of datasets was put forward.In the end,the dual band object detection datasets and object tracking datasets of the dual band made by the scheme were showed,which could provide data support for the research of dual band fusion detection and tracking based on deep learning.