Video target detection method and device and image processing equipment

Abstract

The embodiment of the invention provides a video target detection method and device and image processing equipment. a MaskRCNN network is pre-stored in the equipment, and the equipment comprises a CNNmodel, an RPN, a detection branch and an embedded branch. The image processing device inputs the current image frame into a CNN model for the current image frame of the to-be-detected video to obtaina first feature map; the first feature map is input into an RPN to obtain a plurality of candidate regions, the candidate regions being regions including a to-be-detected target; For each obtained candidate area, the embedded branch generates a first vector corresponding to the candidate area, whether a second vector similar to the first vector exists or not is searched for in a stored embedded list, and if yes, the first vector is updated; And the detection branch predicts the position of the candidate region by taking the updated first vector as time sequence information.

Description

Technical field [0001] This application relates to the field of image processing technology, and specifically to a video target detection method, device and image processing equipment. Background technique [0002] Object detection has been developed for many years as a classic research field in computer vision, but so far, most of the research has focused on object detection in static images. There are relatively few researches on object detection in video images. [0003] At present, the research on object detection in video images is mainly focused on adding timing information to the network in the following two ways: First, use optical flow to warp the detection result of the previous frame (affine transformation mode and distortion) to the current In the network, it is used as the fourth channel input other than the RGB channel; second, when doing Non-Maximum Suppression (NMS), consider the candidate regions of the previous frames. [0004] However, the above two methods are u...

AI Technical Summary

Problems solved by technology

[0004] However, the above two methods are unstable in practical applications

Images