A trajectory optimization based target tracking method and system

Abstract

The present invention provides a target tracking method and system based on trajectory optimization. The target model is constructed in the first frame, including sampling the target sample first, and then marking the sample according to the coverage rate of the target, and extracting the multi-fusion feature training structure. Then, several candidates are taken in subsequent frame pictures, and each candidate training model is used to select the candidate with the highest confidence in the next frame. After multiple frames are iterated in this way, the best short trajectory of the target is selected according to the average confidence of multiple trajectories formed in consecutive multiple frames. Therefore, by introducing the timing stability of the target in consecutive frames, the present invention overcomes the problem that the traditional static detection method has insufficient model discrimination due to ignoring the spatio-temporal information, so that the best target cannot be selected and the model drifts, and proposes a A structured output classification model based on multi-feature fusion can effectively improve the discriminative power of the model.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to an object tracking method and system based on trajectory optimization. Background technique [0002] The goal of object tracking is to estimate the motion state and trajectory of the object of interest in a continuous image or video sequence. It is an important branch of computer vision and is widely used in various fields such as behavior analysis, video surveillance, intelligent transportation, and national defense construction. In recent years, although object tracking has made great progress [1-5] , but due to environmental factors such as illumination changes, scale changes, and local occlusions, the robustness and stability of current tracking algorithms still have certain limitations. [0003] Existing target tracking algorithms can be roughly classified into generative methods and discriminative methods. The generative method refers to finding the most simila...

AI Technical Summary

Problems solved by technology

In the process of building a classifier, if you simply assign binary labels to the training samples, it will bring ambiguity on the labels, and the model will also greatly reduce the discriminative power due to the introduction of wrongly labeled samples.

Images