A tracking method and system based on a convolutional neural network and a Bayesian filter

Abstract

The invention provides a tracking method and system based on a convolutional neural network and a Bayesian filter. The method comprises the steps of based on preset training sets, performing pre-training on a convolutional neural network to obtain a primary model of the convolutional neural network; receiving a video stream with a tracking target input by a user, performing tracking on the tracking target in the video stream via the primary model and performing fine adjustment on the parameters of the primary model by using the fine adjustment technology to obtain the final model of the convolutional neural network; receiving a monitoring video stream with the tracking target input by the user, performing identification and tracking on the tracking target in the monitoring video stream automatically via a post-substitution TLD algorithm and updating a target model set and a background set via a Bayesian filter. By training a convolutional neural network to generate a final model, a tracking target can be identified automatically in a monitoring video stream; samples can be updated by using a Bayesian filter, so that long time tracking of targets is realized and the operation experience of users is improved.

Description

technical field [0001] The present invention relates to the technical field of tracking, and more specifically, relates to a tracking method and system integrating a convolutional neural network and a Bayesian filter. Background technique [0002] Most of the existing research on TLD (Tracking-Learning-Detection, target tracking algorithm) is based on manually identifying the target to be tracked, but in the real-time tracking system of actual production and life, we will find that such an operation method is not practical. For example, in traffic and factory automation production lines, the time when the target appears in the surveillance video stream is uncertain. If the user is required to manually mark the target to be tracked, the user needs to manually mark the target to be tracked every time the program is initialized. This user experience is poor, because in the process of manually marking the target to be tracked, the target to be tracked may disappear from the sur...

AI Technical Summary

Problems solved by technology

For different tracking objects, in order to achieve a good tracking effect, different features should be used to describe the image block to be detected, so that the detector can work better, which involves the problem of feature selection, so from the technical realization In terms of this, it is relatively difficult for programming to realize

Images