Vision-based Implementation Of Motion Target Detection Algorithm For Complex Scenes

This paper takes the infantry robot vision system of RoboMaster,a global high-level robot competition,as the research object.The object of study is in a complex environment,and the existing vision detection algorithms have problems of unstable recognition,low detection accuracy,and long operation time for detecting moving targets.Through the algorithm research and design of the target detection of the vision system of the infantry robot,we aim to assist it in achieving the function of automatically targeting enemy units in a stable,accurate and fast manner.The main work of this paper is as follows.First,it is proposed to evaluate the quality of motion images in complex scenes in terms of luminance,blur,and color shift.Since the recognition target is in a complex environment such as background,terrain,and motion state,the image will have different degrees of degradation problems in terms of brightness,sharpness,and color,etc.Therefore,this paper designs three indicators to evaluate and filter the image quality,and passes the validity verification on the image database.Second,a multi-feature fusion target detection algorithm is designed around identifying the features of target objects,and the task is divided into three parts:image segmentation,light bar screening pairing,and character classification.First,a two-stage image segmentation method with global pre-segmentation and local adaptive segmentation is proposed to address the problem that it is difficult to accurately segment the region of interest for moving targets in complex environments.In the experimental comparison,the recall rate of the segmentation method proposed in this paper is 99.22% and the accuracy rate is 98.90%,which proves the effectiveness of the method in this paper.Then,for the interference problem of similar color and shape regions,the machine learning method of random forest is used for screening.By analyzing the physical geometric features of the recognition objects,six feature variables are identified as the conditions for light bar pairing and armor plate screening.Compared with other classical methods,the random forest-based screening method has the best performance with a recall rate of 98.37%and an accuracy rate of 98.25%.Finally,in order to distinguish the troop types,the ROI of the target was trained and experimentally verified using the SVM+HOG method,which performed with a recall and accuracy greater than 97.14% on different characters.Third,the implementation of the visual detection algorithm.In order to implement the hitting strategy of the vision system in the competition,the correspondence of the target under the four major coordinate systems is analyzed,the internal and external parameters of the camera are measured by camera calibration,and the relative poses of the target are solved using the PnP algorithm.Then,we combine the target’s character information,attitude and distance for autonomous decision making,and provide the correct command information to realize the targeting function to the enemy units.Finally,it is experimentally verified that the visual detection algorithm designed in this paper can achieve stable,accurate and fast recognition of moving targets in complex scenes.