Research On Incremental Learning Method For Intelligent Robot Target Detection With Regularization Constraint

Object detection has grown in importance as computer vision technology has advanced,and it is now utilized extensively in sectors like intelligent robots,autonomous vehicles,and other areas.Traditional target detection models,on the other hand,update their own model parameters to adapt to the new task when learning a new task in an unfamiliar environment.This causes a catastrophic forgetting problem,which can be effectively solved by fusing incremental learning with target detection models.Existing incremental target detection methods rely heavily on knowledge distillation to constrain the model’s parameter updates during training in an effort to maintain a balance between the model’s performance on old and new tasks.However,improper settings of distillation constraints in the model can result in issues such as excessive constraints and uneven distribution of distillation outputs,which can negatively impact the model’s final performance.This paper investigates the problems of distillation over-constraint and uneven distribution of distillation output in the incremental target detection model with distillation constraint method,and deploys the research results on the intelligent robot platform for real-world scene verification.The primary research for this paper is as follows:(1).In order to address the problem of over-constraint in knowledge distillation,this paper proposes a regularization constraint method for incremental target detection with empirical replay.The image repository is configured to store the image data about the old and new classes at the level of training data,and the image data about the old classes is used in the model distillation process at a predetermined iteration interval to improve the efficiency of knowledge transfer and alleviate the constraint of over-distillation.Concurrently,the new image data are used for meta-learning to modify the model to the gradient direction of the new task and improve the model’s performance in the new task.(2).In order to resolve the anomalous distribution of distillation output data,this paper proposes a regularization constraint method for incremental target detection based on multi-network mean distillation.The output data are zero-averaged prior to training the model so that their distribution is always centered on 0.The processed results are then incorporated into the distillation loss calculation in order to reduce the disparity in distribution between the output data of the teacher network and the output data of the student network,which is the result of knowledge distillation.In addition,for the Faster RCNN model with multi-network structure,this paper enhanced the distillation output of the input terminal and output segment and adopted the constraint mode of adaptive distillation for RPN in the middle segment to extract more information about old classes in the middle layer and improve the model’s robustness.(3).Based on the research conducted in the first two sections,this paper conducts incremental learning application experiments in real-world contexts using the intelligent robot platform.On the intelligent robot,the trained model is deployed,and two modes of image detection and real-time video detection are configured to perform incremental detection and recognition of the actual scene under varying illumination and in a complex environment.The results show that the intelligent robot can achieve optimal recognition in bright light and in a single environment,and the results will also provide a basis for the development of subsequent intelligent robots with incremental learning capabilities.