Research On Pedestrian Detection Algorithm In Complex Scene

Pedestrian detection is an important research direction in the field of computer vision,providing the foundation and technical support for research in multiple visual fields,such as pedestrian re-identification,pedestrian tracking,and human pose analysis.It has been widely applied in various domains,including unmanned driving and intelligent transportation.In the past decade,the development of deep learning in visual detection tasks has greatly improved the performance of pedestrian detectors.However,pedestrian detection remains a highly challenging task in crowded and complex scenes.In scenarios with dense pedestrian flows,such as commercial streets,airports,and supermarkets,there are numerous occlusions and interlaceent between pedestrian targets and non-targets,which makes it difficult to distinguish the boundary of target instances and affects the performance of the detector.Furthermore,the various manually designed components in existing detectors significantly limit the universality of the detection model in practical tasks,further reducing the actual performance of the detector.In order to solve the pedestrian detection dilemma in complex scenes mentioned above,this thesis proposes an extended multi-level positioning detection method.This method is based on a predictive training basic model that is more focused on complex scene detection tasks.It combines the Anchor free algorithm with the Anchor base algorithm.On the basis of completely abandoning the manually designed components,the detector can more accurately detect Pedestrian target,the main research content is as follows:(1)A pre-training base model focused on crowded scenes.To adapt to the needs of object detection tasks in crowded scenes,a standard pre-training dataset for crowded scenes was constructed by combining four different degrees of occlusion datasets,enabling the pre-training model to learn more robust features and better adapt to complex scenes with crowded crowds.In addition,based on the Image Net pre-training model,corresponding weight optimization strategies were proposed for classification and detection tasks to further improve the accuracy and generalization ability of the model.Finally,a pre-training model focused on object detection tasks in crowded scenes was trained,improving the actual performance of the model.(2)Diffusion-based multi-level localization detection network.A novel Anchor-free detector based on Gaussian kernel prediction is introduced as the first stage of the entire detection network.The coarse detection of this detector provides more accurate candidate proposal boxes and various prior information for the subsequent detection network,while effectively predicting some lightly occluded and unobstructed instances,greatly reducing the computational overhead of the subsequent network.Then,based on the R-CNN network,a self-supervised detection head is proposed,which takes the center point obtained by the first-level network as the diffusion base point,and uses the Gaussian kernel center region as the supervisory clue for predicting whether the proposal box is occluded or suppressed,assisting the first-level network to predict more accurate object instance prediction boxes.(3)A adaptive post-processing algorithm based on confidence optimization.The algorithm uses the prediction results of the first-level sub-network to gradually enhance the setting degree of the real instance,and weaken the setting degree of the registered remaining frames,so as to obtain high-quality candidates with high consistency between the classification setting and the positioning accuracy.Finally,according to the crowd confidential information obtained from the predicted detection results,an appropriate selection threshold is selected for the Non-suppression algorithm.Through this dynamic lottery adaptive mechanism,the best detection effect is achieved in different density crowds.Finally,the proposed method is extensively experimented on two challenging pedestrian detection benchmarks,including City Persons and Crowd Human datasets.Compared with other state-of-the-art methods,the method proposed in this thesis achieves remarkable performance on the challenging crowded dataset,which strongly verifies the effectiveness of the method in this thesis for detecting pedestrian objects in dense scenes.