Research On Pedestrian Detection Based On Deep Learning Methods

Pedestrian detection aims to analyze whether there are pedestrians in the pictures or videos and give their locations,which is usually used in intelligent monitoring,intelligent robots,and autonomous driving.As an important pre-processing technique of many tasks,it is often combined with pedestrian tracking,person re-identification,pedestrian analysis,and other technologies.Consequently,the performance of the pedestrian detection algorithm closely affects the effectiveness of subsequent tasks.It is necessary to improve the accuracy of the pedestrian detection algorithm.Although pedestrian detection has made great progress in recent years,there are still some problems need to be solved,such as severe occlusion,vast scale differences,complex backgrounds,and scene transfer.We focus on two of the key problems:(1)The distance between the pedestrian and the camera directly influences the size of the target in the image,which causes a significant problem of the scale difference.It is a key factor which affects the performance of the detector.Thus it is important to design a pedestrian detector with less sensitive to scale.(2)There are many different real-world scenarios.It is difficult to label the data in all scenarios,and what we can obtain is unlabeled data more.It is worth transferring the model trained with labeled data to the scene without labeled data.In this paper,we conduct the researches on these two problems,the main work is as follows:In the first work,we focus on multi-scale feature extraction and aggregation.Multiscale feature representation is a common strategy to handle the scale variation in computer vision tasks.Existing methods simply utilize the convolutional pyramidal features for multi-scale representation,resulting in limited performance improvement in multi-scale pedestrian detection.We first analyze the differences of the features in different scale.Then we proposed a bidirectional feature enhancement module,which utilizes the complementarity between high-level features and low-level features to enhance the features in different scale.In order to extract multi-scale pedestrian features of different layers,we design a prior-based multi-scale feature extraction network for the pedestrian,which focuses better on the pedestrian area.Finally,we apply the adaptive multi-scale feature fusion method to fuse the features of different layers into the final predicted feature map,which makes our model more tolerant of multi-scale targets.In the second work,we focus on unsupervised domain adaptation in pedestrian detection.Since there are many different scenes,it is tedious to label the data in every scenarios for training model from scratch.We hope to transfer the supervised model trained on the source data to the unlabeled target data by transfer learning,making a better performance on the target data.We directly simulate the unsupervised domain adaptation problem as semi-supervised learning,and develop the Mean Teacher method to transfer the model on source data to the target data.By pursuing the consistent feature representation of the teacher model and the student model,the teacher model and the student model can be promoted to learn together to improve the detection performance in unsupervised scene.