Real-time target tracking and detection method and system based on fully convolutional neural network

Abstract

The invention discloses a method and system for real-time target tracking and detection based on a fully convolutional neural network. The method includes: S1, performing data enhancement processing to obtain training samples; S2, combining the training samples and the target segmentation map corresponding to the first frame with Combine the color channel dimension; S3, combine the target segmentation map corresponding to the second frame of the training sample and its transpose map in the color channel dimension; S4, construct a full convolutional neural network, which consists of a full convolutional network and a discriminator network Composition; S5, train the discriminator to judge whether the segmented image is fake data or real data generated by the full convolutional network; S6, use cross-entropy to calculate the loss value 1 and the loss value 2 of the segmented image and its label; S7, combine steps S5 and S6 is performed in turn until the fully convolutional network generates as close as possible to the real human-drawn object segmentation map. The present invention relies on less data, has fast calculation speed, and has immediacy, and can track the target in the video while shooting the video.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and relates to a real-time target tracking and detection method and system based on a fully convolutional neural network. Background technique [0002] Object tracking and detection has always been the focus of research in the field of image and video detection. It plays an important role in traffic image processing, crux tracking on medical images, and special video processing (such as mosaicing). [0003] Commonly used object tracking and detection methods include Struck, SCM, ASLA, KCF, TLD, and DCF, etc., but these methods have many shortcomings, such as low recognition accuracy for some, and only for specific categories of objects. When it is necessary to identify new When it comes to object types, the algorithm needs to be rewritten, which makes its development cycle longer, and most current filtering detection-based algorithms cannot achieve high semantic segmentation accuracy. [0...

AI Technical Summary

Problems solved by technology

[0003] Commonly used object tracking and detection methods include Struck, SCM, ASLA, KCF, TLD, and DCF, etc., but these methods have many shortcomings, such as low recognition accuracy for some, and only for specific categories of objects. When it is necessary to identify new For different types of objects, the algorithm needs to be rewritten, which makes its development cycle longer, and most current algorithms based on filter detection cannot achieve high semantic segmentation accuracy

[0004] In recent years, target tracking and detection algorithms based on deep learning have emerged. Although these algorithms have higher performance, they also have slow computing speed, difficult training, and strict requirements on the size of the input image.

In practical applications, there is often a need for multi-target detection. The general method can only track a single target. When multi-target tracking is required, it is necessary to use multi-thread synchronization to track each unit separately, or introduce other methods. Computing speed has a very bad effect

Images