Research On Pedestrian Detection Algorithm Based On Deep Learning

With the rapid increase of car ownership in the city,road traffic safety has become an important issue that cannot be ignored.The Advanced Driver Assistance Systems(ADAS)has emerged as the times require.ADAS refers to equipment that uses advanced technologies such as automotive sensors and computer vision to reduce accident rates.As an important part of ADAS,the pedestrian detection system can timely report the road condition information to the driver to achieve the purpose of reminding the driver to drive safely.Due to the large number of road traffic participants,the situation is usually complicated.How to design an efficient and safe pedestrian detection system to help drivers drive safely becomes a challenging topic.For pedestrian detection algorithms,accuracy and speed are two of the most important requirements.The Accuracy ensures that reduce missed detection and false detection of pedestrians.The speed ensures that the system can quickly detect pedestrians on the road and alert the driver to respond as early as possible to avoid traffic accidents.Due to the complexity of the road environment,there are still some problems in the detection accuracy and real-time performance of the pedestrian detection algorithm,which is difficult to be suitable for practical applications.In response to the above problems,this paper has done the following work:First of all,aiming at the problem that the percentage of pedestrians in some natural scenes is small(hereinafter referred to as small target),the extracted features are easily lost,and the detection accuracy is low,a pedestrian detection method based on candidate region and parallel convolutional neural network is proposed.First,for the candidate region extraction section,Selective Search is improved to make it more suitable for pedestrians in this category of candidate region extraction;then,using Edge Boxes to filter a large number of pre-candidate regions extracted by Selective Search,and finally a small number of high-quality candidate regions are obtained.When using Convolutional Neural Network(CNN)extract feature,deeper convolutional neural networks can extract richer and more abstract high-level features,but at the same time,the small objects can easily cause feature loss,adding shallow convolutional neural network to build a Parallel Convolutional Neural Network(PCNN)to extract deep and shallow feature inputs into the fully connected layer using classifier classification.Finally,the proposed method is applied to pedestrian detection.The experimental results show that the proposed method can improve the detection accuracy of small target.Secondly,in view of the slow speed of pedestrian detection algorithm,this paper adds a coordinate mapping method based on feature extraction using CNN.By inputting the entire image into the CNN instead of hundreds of candidate regions into the CNN,only one feature extraction is required for the entire image,and then the candidate region on the feature map is extracted using the coordinate mapping method at the last convolutional layer.By using coordinate mapping,an image is changed from thousands of feature extractions to one time,which not only avoids the extraction of a large number of repeated features,but also greatly improves the speed by greatly reducing the number of feature extractions.At the same time,because the existing coordinate mapping method is easy to cause the detection accuracy to drop,this paper improves it,so that the algorithm can improve the detection speed while maintaining the detection accuracy.Finally,we completed the interface design and function implementation of a pedestrian detection system.Through the interface operation,the pedestrian detection steps can be implemented more conveniently.