The SSD Vehicle Detection Research Based On Context And Spatial Attention

With the rapid development of the national economy,the number of car ownership continues to increase,and the pressure on traffic supervision faced by society is increasing too.A series of urban traffic problems have followed,resulting in country and individual’s loss with different levels of human,material and financial resources.In recent years,with the rise of artificial intelligence and big data technology,in order to alleviate and solve the urban traffic problems and negative impacts caused by them in a timely manner,China has proposed to vigorously promote the intelligentization of traffic supervision construction,and proposed new demands and challenges for intelligent transportation systems and researchers.Vehicle detection is one of the key technologies and the basis for some other important technologies,such as traffic flow statistics,vehicle tracking and calculation of distance between the cars.In real-world applications,vehicle detection requires real-time detection of vehicles present in images or video frames and the precise location of the vehicle.Therefore,it is of great importance and significance to study a vehicle detection algorithm with better performance.The traditional vehicle detection algorithms are mainly divided into modeling based front and back scene separation and machine learning based manual feature extraction.The former--such as interframe difference method,optical flow method,background difference method,etc.The latter like HOG manual feature extraction and DPM algorithm.At present,the detection technology based on deep learning has developed rapidly,which is a research hotspot of machine learning and has achieved good results in many scenarios.This thesis focuses on the target detection algorithm based on deep learning and analyzes it.The SSD model with good performance in accuracy and speed is selected as the basic model.An improved model based on context and spatial attention is proposed and achieving vehicle detection by the model.The main contents and innovations of this paper are listed as follows:1.According to the feature information of shallow network in SSD model to predict the target object will lack the deep semantic feature information.This paper is inspired by context information fusion in DSSD model.Through experiments,two specific convolution layers in SSD are selected and do context fusion between them.Through context information fusion,enables the SSD model to compensate for the expression of shallow feature information and improve detection accuracy.The experimental results show that the performance of the model with the introduction of context on the datasets VOC2007 and VOC2012 is better than the model without introducing context,which increased the precision by about 1.5%.2.In view of the possible deformation of different vehicle detection data images and the loss of certain location feature information in the pooling layer of the convolutional neural network.This paper introduces the spatial transformation network based on the SSD model adding context information fusion.The localisation network generates transformation parameters for the data enhancement in the SSD model training process,further enhances the spatial invariance,and compares the performance of the vehicle detection with other three algorithms.The experimental results show that the performance of the improved model is still better.