Research On Pedestrian Detection In Mixed Traffic Scenarios Based On Deep Learning

Pedestrian detection is one of the important components of computer vision,which is widely used in intelligent driving,assisted driving and intelligent video surveillance.In the context of mixed traffic in China,pedestrians in the scene have multi-targets,obstructions,and complex background interference.It is difficult for traditional pedestrian detection methods to obtain good detection results.In response to these problems,this paper combines the deep learning-based pedestrian detection method with this special background,and optimizes the detection effect from multiple aspects.For the pedestrian training set in the selected mixed traffic scene,the training difficulty of the network model is reduced from the aspects of diluting the impact of the traffic background and making the pedestrian characteristics more obvious.Pre-process the training set sample data,perform initial correction and gray level conversion on the training set images,and then perform image enhancement processing on the training set images to reduce the difficulty of extracting pedestrian features in the scene and enhance the detection network The effect of training.Aiming at the recognition and detection in mixed traffic scenarios,in order to meet the requirements of accuracy and speed,this paper builds a pedestrian detection network model using YOLOv3 as the prototype network.In order to achieve better accuracy,this paper improved the feature pyramid network(Feature Pyramid Network,FPN)multi-scale feature fusion method.The scale conversion method in Scale-Transferrable Object Detection(STDN)is used to replace the interpolation upsampling method in FPN,and the method of multi-scale feature fusion in this paper is designed and implemented.The Darknet-53 network in YOLOv3 is selected on the prototype network to meet the requirements of real-time monitoring.Using the designed multi-scale feature fusion method to improve the Darknet-53 network,a pedestrian detection network in a mixed traffic scene is constructed.And in the comparison experiment with the original YOLOv3 algorithm network,it has achieved excellent detection results.In order to achieve better detection results and avoid gradient explosion and loss divergence during training,soft sampling is used to reduce the impact of simple negative samples during training.In terms of loss,since the detection target in this paper is relatively single,the classification loss in the detection process is reduced to improve the network training effect.The pedestrian detection network model constructed in this paper is used to conduct detection experiments in actual traffic scenarios.In order to verify the practical application value of the detection network model,the test set is all from actual pictures that meet the requirements.From the analysis of the detection experiment results,the pedestrian detection effect in the mixed traffic scene is outstanding,which demonstrates the research value of this article.