Research On Network Model And Algorithm Of Video Object Detection Based On Anchor-free

Object detection is an important branch in the field of computer vision.Its task is to determine the object we need to detect in the image.Usually a rectangular box is used to frame the specific location and size of the object in the image,and to determine the object category for output.At present,there are many representative algorithms for image object detection.However,the detection of video objects is progressing slowly.Although video data is more prone to abnormal deformation or occlusion problems than image data,the video data also contains more time-series information than image data.Common object detection algorithms can be divided into two types,one is two-stage detection,and the other is single-stage detection.The former first obtains a candidate box(proposal),and then performs classification and regression on this candidate box,which has high accuracy.The latter uses only one network to directly detect the object and has a faster speed.The Center Net algorithm belongs to the standard Anchor-free algorithm.It generates a heatmap through a Gaussian kernel,predicts the center point of the target object,and draws detection box through the center point and the predicted length and width.This article is to use Center Net object detection algorithm as the basic detector to improve the video object detection.First,the CBAM attention mechanism is added to the backbone network to improve the detection of a single frame image,then use the heatmap information of adjacent frames to perform feature fusion on the current key frame heatmap,Improve the detection effect of key frame.Through experiments,the algorithm researched in this topic can identify and detect multiple categories in the detection task.Through adjustment of parameters and multiple iterations of training,it is also 3% m AP higher than the traditional Center Net algorithm on the self-selected VID dataset.