Research On One-Stage Object Detection And Cross-Dataset Training

Object Detection is an important subject in the field of computer vision,which plays an important role in the application of automatic driving,security and robot.Current popular object detection algorithms are mainly base on re-gion proposal methods using deep learning,which are divided into two main series,one-stage and multi-stage.The one-stage algorithm has the advantage of fast speed,but its accuracy is not very high.The multi-stage accuracy has advantages,but its speed is slow,and it is often not real-time in practical use.At the same time,in the actual algorithm,the current mainstream algorithms can-not solve the extra labeling problem caused by the increase in classes in actual use.Facing the growing classes demand in practical applications,a lot of re-peated labeling of the original data is required.In order to solve these problems,this paper proposes some practical and effective solutions.This paper first proposes some improved algorithms for the one-stage de-tection algorithm of object detection,which improves accuracy while ensuring real-time speed.We also analyze the differences between the anchor-base and anchor-free algorithm,and propose a strategy for adaptively dividing positive and negative samples.Finally,in order to solve the problem of category growth in practical application,a new algorithm is proposed.Aiming at the problem of fixed single-level feature receptive field in one-stage detection algorithm,the concept of dense receptive field is proposed,which combines the receptive fields of different feature layers to better detect the object.Aiming at the differ-ences between the positive and negative samples based on the anchor-base and the anchor-free algorithm,an adaptive positive and negative sampling strategy is proposed,and the threshold of the positive and negative sample attributes is dynamically set.In order to solve the class growing problem often encountered in practical applications,a cross-dataset training method is proposed,which only needs to label the data of the growth classed,which solves the problem of repeated labeling of existing data.Finally,through the design of reasonable and effective comparative experiments,very good results have been achieved on various authoritative public datasets,and the effectiveness of the algorithm proposed by this paper is verified.