Research On Deep Video Object Detection Based On FlowNet

Object detection is a very popular research direction of Computer Vision.So far,there are several representative algorithms for single image detection in academia.But the research progress in video object detection is in a relatively backwater.Comparing image data with video data,the latter has a larger amount of information and is more complicated.Nowadays,Internet technology and mobile communication technology are developing rapidly.Whether using computer or mobile,we can see video data plays a more and more important role in our lives.The task of object detection is to find all the interesting objects of the input image.The task of video object detection is to find all the interesting objects in the video,and output their exact positions,sizes and classifications.However,in a sequence of video frames,adjacent frames have great similarity,and there is a contextual relationship between the front and back frames.This is a feature that single images do not have.Therefore,it seems that video object detection is similar to image object detection,the actual algorithm design is much more different.This topic is based on the previous MSRA's fixed interval key frame selection algorithm and dense feature enhancement algorithm for optical flow detection of video objects.We proposed an algorithm based on fragmented key frame insertion strategy.The new algorithm is still based on feature-level detection principle in the overall structure.That is to say,it combines the feature extraction network and the region-based fully convolution network.The main idea of the algorithm is to select key frames in a segmented way using a sequence of key frames at regular intervals as the initial state.First,determine whether the similarity between the current key frame and the next key frame reaches the similarity threshold.Second,insert a new key frame,otherwise skip to the next key frame to continue the decision.Until all key frames have been determined,a sequence of key frames with a distribution trend of fragments is finally generated.Through experiments,some parameters of this algorithm can be adjusted wo get to different scenarios.