Technical Research On Pedestrian Detection Based On Vision

With the development of technology and the arrival of the era of artificial intelligence,there is a wave of artificial intelligence in all walks of life.In the intelligent transportation system,how to detect pedestrians in the traffic environment becomes the main focus.Due to the traffic environment is very complex,everyone is the participants of the traffic in the strict sense and is involved in transportation in the form of different everyday,traffic safety is everybody's life on the road,which has certain requirements for the detection accuracy and speed of the detection algorithm on the road.The traditional pedestrian detection method can meet certain requirements in speed,but there are many deficiencies in the detection accuracy.With the in-depth study of pedestrian detection,the data set for pedestrian detection has been greatly improved in both quantity and quality through continuous replenishment and optimization.Although new method based on convolution neural network has improve on the accuracy of bigger,the new problem is that the amount of required computation will be increased with the addition of the amount of required data.To some extent,the development of parallel computing in the current environment has alleviated the problem.Firstly,this paper introduces many problems and factors in pedestrian detection,such as the public data set,feature extraction method and classifier selection method for pedestrian detection.There are also regional recommendations to replace traditional pyramid sliding window methods and non-maximum suppression algorithms(NMS)for reducing overlapping detection boxes.Through concrete study found that the relationship between the target detection and segmentation is very tight,they can realize each other optimization,each other between them that is can into the split in their thoughts in order to improve efficiency,in other words,pedestrians detection can incorporate the idea of segmentation into the test to improve efficiency.Secondly,we use Mask RCNN general framework to detect the images in the test set.In this paper,we detail the design approach that is pedestrian detection network of based on Mask RCNN,such as feature extraction,regional network recommend network,the classification of the candidate window and segmentation.We use MS COCO data sets to train the network,and it was tested in the latest pedestrian detection data sets that is Cityscapes data set.The experimental results show that this method can achieve high efficiency and accurate in pedestrian detection.Aiming at the problem of pedestrians small target leak,we did a systemic analysis.We did preliminary validation about the problem that the distribution of the data set target dimension will affect the target detection performance of the model in different scales,but changing the overall distribution of data set is relatively difficult.So the adjustment of the input image can help to solve this problem.This can affect target dimension distribution of test set,and solve the problem of small targets were missing.Combined with the advantages of Mask RCNN's target segmentation accuracy,an optimization algorithm is designed to correct the situation of the driver and passenger in the car.Finally,using the method of feature fusion,we make the Mask RCN combined with the feature of image segmentation.This network of image segmentation uses DeepLabv3.After fusing image segmentation network,the accuracy of medium and large targets improved.