Target detection method for self-adaptive non-maximum suppression

Abstract

The invention discloses a target detection method for self-adaptive non-maximum suppression. The method comprises the steps of S1, selecting an initial candidate box set for iterative processing so asto traverse and rank and score candidate boxes in the initial candidate box set, and composing all candidate boxes which do not get highest scores in ranking and scoring into a remaining candidate box set; S2, obtaining adjacent target region graduation of two adjacent candidate boxes based on difference of attention maps of the two adjacent candidate boxes in the remaining candidate box set; S3,constructing a self-adaptive score attenuation function based on the adjacent target region graduation of the two adjacent candidate boxes, and automatically endowing an attenuation coefficient corresponding to scores of the two adjacent candidate boxes based on a calculation result of the self-adaptive score attenuation function; S4, rescoring the two adjacent candidate boxes and discarding thecandidate boxes whose scores are lower than a threshold value; and S5, iteratively repeating the steps S2-S4, judging whether the number of the candidate boxes in the remaining candidate box set is 1,if so, terminating target detection, and outputting a final candidate box fusion result.

Description

technical field [0001] The invention relates to the technical field of target detection, in particular to an adaptive non-maximum suppression target detection method. Background technique [0002] In practical applications, object detection often needs to deal with various scenes, especially complex scenes. For example, in urban environments, object detection systems often need to face scenes that contain a large number of objects and these objects overlap each other. The object detection task poses a great challenge. In this case, object detection often leads to confusing detection results due to multiple candidate regions regressing in the same region of interest, and non-maximum suppression (NMS) is usually used as a post-processing step to obtain the final processing results. A major problem with non-maximum suppression is that it sets the score of adjacent candidate boxes that exceed the NMS threshold to 0, that is, makes a hard decision by deleting the detection resul...

AI Technical Summary

Problems solved by technology

Therefore, when the targets are close to each other, no matter how the threshold is adjusted, traditional non-maximum suppression is doomed to sacrifice precision or recall

Images