A target detection system and method based on adaptive combined filtering and multi-stage detection

Abstract

The invention relates to a target detection method and system based on adaptive combined filtering and multi-level detection, including a target detection module combining a moving target detection unit and a salient target detection unit, a target recognition module based on a convolutional neural network, and a combined decision-making based It cooperates closely with the target tracking module of multi-channel image features; target detection module, target recognition module and target tracking module to form a stable and reliable target detection system. The output of the system is the position information of the detected candidate target, the target category information and the position information of the selected target obtained by target tracking. The target detection system is implemented on a high-performance multi-core DSP chip, and has been optimized for the multi-core DSP chip to realize real-time target detection and target tracking, as well as fast target recognition functions. The invention has strong practicability and high feasibility, and can be conveniently integrated into various solutions requiring target detection to realize intelligent target detection, identification and tracking.

Description

technical field [0001] The present invention relates to the field of pattern recognition and machine learning, in particular to a target detection system and method based on adaptive combined filtering and multi-level detection. Background technique [0002] When human beings use vision to observe the scene, they will first find some interesting objects in the scene, and then judge their categories through perception, and then continue to observe the most interesting objects. In the field of pattern recognition and machine learning, in order to imitate human behavior to realize computer vision, these tasks are abstracted into target detection, target recognition and target tracking. [0003] Target detection is the first step in target detection, that is, to detect targets that need further processing from the scene. Commonly used target detection methods include inter-frame difference, background modeling and methods based on deep neural networks. Among them, the inter-fra...

AI Technical Summary

Problems solved by technology

The features with scale and rotation invariance can finely represent the local area in the image, but the information contained in other areas in the image has not been well utilized

The histogram of gradient statistics adopts a dense representation, which can accurately extract the gradient details of the image, but cannot distinguish images with similar textures well

Color features can distinguish images of different colors, but cannot represent the shape information of images well

The representation forms of the above features are quite different. In the existing methods, only one or two of them can be used at the same time, and the features are not organically combined, so that the algorithm cannot exert the maximum performance of the comprehensive features, and the obtained target tracking method is stable. Not high enough, the target tracking result is not accurate enough

And there are many challenging scenarios in object tracking, the existing method of object tracking through a single tracker is prone to tracking failure during operation

Images