Research On Video-based Target Detection Algorithms

In recent years,artificial intelligence technology has developed rapidly,and computer vision technology in one of its branches has also made breakthroughs.The application of these new technologies has greatly affected people's life.Video object detection is a hot research topic in the field of computer vision.It is widely used in the fields of autonomous driving,intelligent transportation,and security.It is an important content of video information processing and analysis.However,there are still many difficulties in the application of object detection algorithms in the video field.For example,the size of the video object is too small,the objects overlap with each other,the video is jittery,and the weather is bad.This paper studies related video object detection algorithms in complex situations with related theories.Video moving target detection algorithms are mainly divided into two stages:moving target detection and moving target recognition.In the moving object detection stage,aiming at the problems of holes in the foreground of the moving target detected by the traditional inter-frame difference method,and the background model of the Gaussian mixture model is easily affected by the foreground pixels.Based on the traditional inter-frame difference method and Gaussian mixture model,the moving foreground detected by the two are fused as the final detection output to obtain a moving target foreground with a more complete contour.For possible occlusions between moving targets,we use convex hull detection to determine whether occlusion occurs.When occlusion occurs,we use RPN network to locate the occluded targets.In the moving target recognition stage,the features of the moving target are extracted through a convolutional neural network.End-to-end tasks of moving targets classification and coordinate regression tasks are completed,the feature extraction capability of the moving target recognition stage is improved,and the complexity of the multi-target classification and coordinate regression modules is reduced.In some scenarios,such as autonomous driving,embedded devices,where hardware computing power is not strong,storage capacity is not high,and certain real-time performance is required,the single-stage image target detection algorithm based on deep learning is used as the basic framework to improve SSD target detection algorithms.We removed the heavy VGG-16 feature extraction network,used the lightweight network model MobileNetv1 as the backbone network,constructed the FPN feature pyramid,and predicted the detection results at 5 feature levels.The lightweight feature extraction network comes at the cost of a small amountof detection accuracy and significantly reduces the amount of model parameters.The FPN pyramid structure enhances the feature representation capabilities of the feature extraction network.The improved algorithm achieved 32.1% mAP on the MS COCO test set.The detection accuracy surpasses the SSD513 algorithm(mAP = 31.2%)with ResNet101 as the backbone network,and its performance is equivalent to the YOLOv3 algorithm(mAP = 33.0%),which can quickly and efficiently complete video target detection tasks under severe weather conditions.This paper uses traffic monitoring videos as research materials and the vehicles in the videos as research targets to carry out research on video object detection algorithms.The videos are annotated with image annotation tools to construct the dataset,and the model is trained and verified on the public dataset.The algorithm in this paper can quickly and accurately complete the video target detection task in complex scenarios.