Pedestrian Detection In Driverless Scenarios

With the development of artificial intelligence science,autonomous driving has become one of the hot research topics in the world.As an important part of self-driving car environmental perception systems,pedestrian detection also gets more and more attention.This paper researches pedestrian detection algorithms based on deep convolutional neural networks,and conducts relevant research on multi-scale detection,night detection and data augmentation.The main research contents are as follows:1)A pedestrian detection algorithm based on multi-scale classification is proposed.Intuitively,the characteristics of pedestrians of different scales are obviously different.For a single classifier,it is difficult to simultaneously distinguish pedestrians of different scales from the background.Therefore,this paper designs a multi-stream framework,which consists of different classifiers and regressors dealing with large-scale proposals and small-scale proposals respectively.In addition,two auxiliary supervisions are employed to balance the effect of two parts of proposals on the back propagation.The experimental results on the City Persons,Caltech and ETH datasets show that the algorithm achieves improvements to the baseline method,especially on the small scale subset.It is worth mentioning that the algorithm reduces the computational cost while improving performance.2)Two night pedestrian detection algorithms based on local image enhancement is proposed.Most existing pedestrian detection algorithms are for day-time and have poor generalization ability on night-time datasets.At night-time,there are obvious differences in the characteristics of pedestrians wherein different illumination regions.An effective method to solve this problem is to enhance the recognizability of pedestrian appearance characteristics and to reduce the difference between different pedestrians by performing local image enhancement on pedestrian regions.Therefore,this thesis respectively introduces an image enhancement module from the perspective of images and features to help pedestrian detection at night.First,a night pedestrian detection algorithm based on pedestrian re-classification network is proposed.The algorithm crops out the corresponding image regions from the original images according to the detection results of the primary pedestrian detection network,and then uses the local image enhancement to process the image regions,which are sent to the secondary classification module for rescoring.Second,a night pedestrian detection algorithm based on feature transformation is proposed.The algorithm first performs local image enhancement on pedestrian regions according to the annotated pedestrian bounding boxes,and then uses the enhanced images to train the teacher model,which is used to guide the student model to train on the original night-time images,allowing the pedestrian features extracted by the student model be close to the teacher model.Experimental results on the ECP dataset show that the two algorithms proposed in this paper are both superior to the baseline method.3)A pedestrian data augmentation method based on segmentation masks is proposed.There are many background images that do not contain valid pedestrian samples in the pedestrian detection datasets.If these images are used to train the model,the extreme positive-negative samples imbalance will occur in the training phase.If the background image is filtered out in the training phase,the amount and diversity of training data will be decreased.This paper designs a data augmentation method.The method first crops out corresponding pedestrians samples according to the annotated pedestrian segmentation masks,and then re-combines pedestrian samples and background images to generate new images according to the statistical distribution of pedestrian sample positions and heights in the training set.The method increases both the amount and diversity of training data.Experimental results on the City Persons dataset verify the effectiveness of the method.