Research On Object Detection In Teaching Scenes Based On Deep Learning

In recent years,the concept of intelligent classroom has gained a lot of attention.One of the most critical applications is to recognize the visual and semantic understanding for teaching scenes by modeling the targets appearing in the scenes to complete the effective analysis,recognition and expression of such scenes.In order to achieve this,in the first,it is necessary to capture the objects appearing in the scene accurately to obtain position and category information.Object detectors based on CNN are widely used to address this problem to get more accurate performance.However,the task of object detection in the teaching scenes will face some challenges such as dense distribution and serious occlusion of the objects.The performance of the typical detectors is difficult to meet the actual needs.This thesis studies some new target detection frameworks based on current universal detectors to achieve higher detection performance.The main research contents of the thesis are as follows:1.A dataset for teaching scenes is built.This dataset includes ten categories of common objects in teaching scenes with the standard rectangular box labels for each instance.There is a dense distribution of objects in the dataset,which is consistent with the distribution characteristics of common objects in this scene and provides a dataset basis for the development of related scientific research in this realm.2.An indoor person detection algorithm based on the discriminative part selection is proposed.With the method of dividing the proposals into sub-regions and extracting more discriminative parts in the sub-regions to improve the problem of generally poor detection performance of human targets due to occlusion in indoor scenes.The method used in this paper can effectively improve the person detection effect of occludes in indoor scenes.3.A dense object detection algorithm using two semantic-aware branches is proposed.The anchor-based and anchor-free strategies are combined to improve the performance of the detector on dense targets effectively.In the anchor-free branch,key-point prediction based on feature map is used to obtain the position and scale information of the anchor box.And another feature adaptive matching module is used to enhance the detection performance of the branch.Using this method can improve the detection performance of the detector for multiple targets in this dense scene.4.A dense object detection algorithm with adaptive non-maximum suppression is proposed.Non-maximum suppression(NMS)algorithm with fixed threshold is not the optimal solution in different scenes.This thesis uses a network which is sensitive to the scene density to adaptively adjust the threshold of NMS in the post-processing to enhance the adaptability of the detector to different scenes,thereby optimizing the final detection result.