Deep Learning Based Key Frame Detection For Sport Video

Target detection and behavior understanding in video become a hot issue in the field of computer vision.It is widely utilized in human-computer interaction system,behavior Monitoring,auxiliary sports training system and so on.In this thesis,by analyzing weightlifting video,we extract the key poses of athlete to assistant coach to train athletes more professionally.The primary works includes as follows:First,we propose a region of interest classification based key pose extraction algorithm.First,we train a FCN network using weight lifting video frames to extract the foreground and to remove the background interference.Furthermore,the traditional CNN network is fin tuned using the ROI images.Finally,the classification selection strategy is design for key pose detection.The experimental results confirms the efficiency of the proposed algorithms.Secondly,we propose a ROI enhancing algorithm by extracting key object skeleton.First,we train a deepskeleton network using weight lifting video frames to get the skeleton.Because each key pose has unique skeleton,we use deepskeleton network to enhance the region of interest.So that we can improve the accuracy of recognition.After skeleton extraction,we fine-tune the CNN classification network and extract the key poses.Thirdly,we designed a weight lifting video key frame extraction system.It contains display module and key frame extraction module.The display module contains the play of video and the display of key frames,and the key frame extraction module contains the feature extraction of lifting video and key frame extraction.