Research On Pedestrian Detection Based On Deep Learning In Complex Background

Pedestrian detection refers that using the relevant visual techniques to identify the presence as well as the location of pedestrians from the input data.As a subdivision of the object detection field,after years of development,pedestrian detection has now been widely applied in many scenarios such as intelligent security and assisted driving.Driven by deep learning technology,the current pedestrian detection algorithm have achieved good performance on many public data sets.However,with a complex background,the pedestrian detection results of deep-learning based methods are not ideal due to factors such as ambient lighting,background texture and pedestrian clothing.There are remain many problems unresolved.To address these problems,this paper studies the pedestrian detection in complex background based on deep learning technology.The major contributions are listed as below:1.We proposed a complex background pedestrian detection algorithm based on SSD and Inception.It is shown that the SSD model can not extract sufficient feature in complex background.We improved the multi-scale feature extraction network of SSD model by exploiting Inception Block and feature fusion.With the context information being shared among specific feature layers,the pedestrian detection recall rate of the proposed model in complex background is improved.2.Based on DenseNet and Inception,we proposed two variants of SSD model to improve its detection performance in complex background.Aiming at the problem that the base network of SSD model has insufficient feature extraction ability,we modified the structure of SSD by changing the base network and adding sub-modules.It is shown that both variants achieved better detection results in complex background than the original SSD model.3.Based on the idea of Focal Loss,we proposed an adaptive focal loss function that can be more effectively applied to pedestrian detection tasks in complex background.Particularly,we improved the selection mechanism of the balance factor and the focusing parameter of Focal Loss by dynamically adjusting the values during the training process.It is shown that the proposed adaptive focal loss function can better distinguish between the positive and negative samples as well as the easy and difficult samples.The training of pedestrian detection models for complex background can thus be more effective.