Human body tracking algorithm based on attention mechanism and double-flow multi-domain convolutional neural network

Abstract

The invention discloses a human body tracking algorithm based on an attention mechanism and a double-flow multi-domain convolutional neural network. The double-flow multi-domain convolutional neural network tracks a human body target in a double-flow input and multi-flow output mode. The apparent characteristics of the human body target are extracted through the original video sequence, the motion characteristics of the human body target are extracted through the frame difference sequence, and the proportion of the motion amplitude of the human body target in the image is taken as an adaptive fusion parameter to realize the adaptive fusion of the apparent characteristics and the motion characteristics of the target; the algorithm effectively solves the problem that the tracking effect is poor when the shape, the scale, the texture and the like of the human body target to be tracked are changed, and feature optimization and improvement are performed on the tracking target in real time by adopting a mode of offline training shared layer network parameters and online learning independent layer network parameters. It can be known through experimental results that the network can achieve accurate human body target tracking on the premise that real-time tracking is guaranteed.

Description

technical field [0001] The disclosure of the present invention relates to the technical field of computer vision, in particular to a human body tracking algorithm based on an attention mechanism and a two-stream multi-domain convolutional neural network. Background technique [0002] The current target visual tracking is mainly divided into two categories: generative model methods and discriminative model methods. The generative method first needs to model the target area, and finds the area most similar to the model in subsequent frames to achieve position prediction. The discriminative method combines image features with machine learning, takes the target area as a positive sample, and the background area as a negative sample, and uses a classifier to find the optimal area in the entire image. In recent years, related filtering methods and deep learning methods have shown outstanding accuracy and rapidity, and are often used by scholars. However, the processing speed of d...

AI Technical Summary

Problems solved by technology

[0004] At present, in the human target detection and tracking algorithm, there are problems of slow real-time tracking of the target, and poor tracking effect when the target itself changes in shape, scale, texture, etc.

Images