Research And Application Of Video Target Detection Algorithm

Object detection is a research hotspot in the field of image processing and computer vision,and it is the basis of multi-object tracking and object recognition.With the development of convolutional neural networks,the image object detection task has made great progress,and the accuracy and speed have been significantly improved.However,in the fields of auto-driving and monitoring,the current object detection algorithms are insufficient.How to achieve highperformance video object detection algorithm is of great significance in practical applications.Therefore,this paper studies the video object detection and designs related algorithms.The main work and innovations are as follows:(1)In the video object detection process,the temporal and spatial similarity of the video are used to improve the efficiency.At the feature level,based on the mainstream video object detection algorithm,a deformable sampling feature flow algorithm is proposed.Aiming at the problem of single transformation and difficult training in the use of optical flow for feature propagation in this kind of algorithm,this paper uses the deformable convolution to optimize the process of optical flow field transformation feature,with almost no increase in computational complexity.The detection accuracy is improved.(2)In the post-processing stage of video object detection,this paper proposes a frame-by-frame detection frame transmission algorithm for the problems of large computational speed and slow detection speed in the post-processing stages of many mainstream video object detection methods.After the conventional non-maximum suppression,the optimal detection score sequence is selected and the detection frame is transmitted,and the feature propagation strategy of the algorithm is improved.This method not only further improves the speed advantage of the deep feature flow algorithm,but also has a certain improvement in accuracy.(3)In this paper,the video object detection technology is applied to specific tasks.For the actual problem of vehicle automatic driving,the vehicle object detection database LIVA_CAR is built and tested.In this paper,the proposed method is tested on ImageNet VID dataset and LIVA_CAR vehicle detection dataset.The performance is better than the comparison algorithm.It proves that the proposed method can effectively perform video object detection tasks and is also robust.