Detection method for video target based on image significance and characteristic prior model

Abstract

The invention discloses a detection method for a video target based on image significance and a characteristic prior model. The detection method comprises the following steps: training a convolution neural network by inputting a positive sample and a negative sample and taking the trained convolution neural network as the characteristic prior model of a target; extracting a salient region of certain image frame in the image sequence and clustering to obtain a significant candidate box; inputting the significant candidate box into the characteristic prior model and predicting a target positionto realize the detection of the target in the image frame; positioning key points in the significant candidate box by utilizing the characteristic prior model, calculating the key points by utilizingan L-K sparse optical flow method and predicting the motion direction of the target; in the image frame, extracting candidates meeting the motion direction constraint and a confidence coefficient threshold value and taking the candidates as a new positive sample and a new negative sample, inputting and updating the characteristic priori model; repeating the steps for each image frame in the imagesequence to achieve the target detection of the image sequence. The detection method disclosed by the invention has the characteristics of capability of realizing target detection, high antijamming capability, high detection precision and good real-time performance; the engineering application ability of a target detection system is enhanced.

Description

technical field [0001] The invention relates to the field of image processing, in particular to the detection of targets under complex moving backgrounds. A method of combining image saliency and prior features of targets is adopted, which improves the accuracy of target detection and reduces the complexity of the algorithm It meets the real-time requirements of video target detection. Background technique [0002] With the development of autonomous driving technology, drone technology, and security monitoring technology, target detection in video has become a hot spot in the current image processing research field. How to accurately locate the target in a complex motion background and how to quickly locate the target with limited computing resources are two major problems that need to be solved in this field. [0003] Most of the traditional video target detection methods use the target motion information between image frames, such as background difference method, inter-fr...

AI Technical Summary

Problems solved by technology

The optical flow method does not need to know any information of the scene in advance, and can accurately calculate the moving speed of the object, but the calculation of most optical flow methods is quite complicated, the hardware requirements are relatively high, it is not suitable for real-time processing, and it is sensitive to noise and resistant to noise. poor noise performance

Not only that, these traditional methods also have this common defect: they can only detect moving targets, but they are powerless for special static targets in the video.

Although this kind of method solves the defect that the traditional detection method is not high in accuracy and can only target moving targets, there are still obvious deficiencies: (1) Independent target detection for a single image cannot maintain the position of each target in the video (2) The use of convolutional neural networks to extract target features requires a large number of data sets for offline training. Feature extraction depends on the hardware environment of the GPU, which has high computational complexity and high power consumption. It is difficult to build a detection system in On the mobile platform; (3) The real-time performance is not high, and it is difficult to meet the real-time video processing requirements of 30 frames per second

But different from the present invention, this method does not consider the update problem of the target feature model, fully utilizes the model that has been trained for target detection, and the tracking performance for the target is relatively poor

Images