Multi-target tracking method based on coarse-to-fine shielding processing

Abstract

The invention discloses a multi-target tracking method based on coarse-to-fine occlusion processing, and the model construction comprises the following steps: adding an occlusion score prediction on the basis of a JDE model prediction head, thereby completing the complete processing of a non-occlusion target and the coarse processing of an occlusion target; on the basis, taking a pedestrian shielding area after mapping and cutting as a training set of the second-step model for training, and completing accurate detection and apparent feature vector extraction of the shielded pedestrians, thereby realizing the fine processing of the shielded target; integrating results output by the two-step model, and finishing pedestrian tracking by using a data association algorithm. According to the invention, the problem that pedestrians cannot be accurately tracked in a shielding scene in the prior art is solved, and the method can well adapt to public environments of various time periods and various pedestrian densities; the method has a better effect on tracking pedestrians.

Description

technical field [0001] The invention belongs to the field of computer vision and monitoring video analysis, and in particular relates to a multi-target tracking method based on rough-to-fine occlusion processing. Background technique [0002] Multi-object tracking is an important part of surveillance video analysis. It can not only be directly used for object trajectory analysis, but also can be used as the research basis for advanced tasks such as object action recognition and behavior analysis. [0003] In order to complete the multi-target tracking task, many mainstream deep learning algorithms have proposed tracking strategies based on detection. These methods split multi-object tracking into a detection module and an embedding module. The detection module completes target detection, and the embedding module uses related algorithms to extract the features of the target. However, there may be multiple repeated calculations between these two modules, which affects the ru...

AI Technical Summary

Problems solved by technology

For the problem of occluded target detection, some improved non-maximum suppression algorithms have been proposed, such as soft-NMS, softer-NMS, adaptive-NMS, etc., and some improved loss functions have been proposed, such as repulsion loss, etc., but these methods all rely on The prediction results of the original detection network, for some scenes, the effect of occlusion detection is not good

Images