Target detection method based on iterative design

Abstract

The invention discloses a target detection method based on iterative design. The method solves the problems of low detection precision and low recall rate caused by adopting post-processing methods such as non-maximum suppression in a traditional target detection method, and is characterized in that a target detection main network and a historical information feature extraction sub-network are constructed; feature fusion is carried out through a fusion mechanism, and training of a target detection model is carried out; and target detection in a scene is carried out through the target detection model after training convergence. The target detection method based on iterative design makes detection precision and robustness effectively enhanced, reduces the difficulty and complexity of training and is more universal.

Description

technical field [0001] The invention relates to a target detection method, in particular to a target detection method based on iterative design. Background technique [0002] In daily life, there will be many scenes where the targets that need to be detected are occluded from each other, such as the detection system of pedestrians in video surveillance in public places, the monitoring system of vehicle flow in road traffic management, and the automatic detection of dense cells under the microscope in medical laboratories system and so on. The overlapping and occlusion between the targets that need to be detected in the above scenarios will pose a huge challenge to the target detection algorithm. How to accurately detect the target in the target occlusion scene is a subject of great concern in both academia and industry, but the detection accuracy and recall rate of the current general target detection algorithm in the target occlusion scene are difficult to meet the require...

AI Technical Summary

Problems solved by technology

How to accurately detect the target in the target occlusion scene is a topic of great concern in both academia and industry, but the detection accuracy and recall rate of the current general target detection algorithm in the target occlusion scene are difficult to meet the requirements

[0003] General target detection algorithms cannot achieve satisfactory results in complex scenes such as target occlusion. The fundamental reasons are as follows: (1) Traditional post-processing algorithms such as non-maximum suppression cannot distinguish Whether the adjacent detection frames are redundant detection frames of the same target or detection frames belonging to two overlapping targets

(2) Moving targets will not only deform but also occlude each other. Traditional target detection models cannot adapt to target deformation and deal with insufficient target information.

Images