With the popularization of intelligent transportation system,the demand of vision-based vehicle detection algorithm and performance become more and more severe.Due to the development of deep learning,convolutional-neural-network(CNN)-based vehicle detection approaches achieve incredible success in recent years.However,existing CNN-based algorithms suffer from the problem that the convolutional features are scale-sensitive in object detection task but it is common that traffic images and videos contain vehicles with a large variance of scales.Besides,training CNN needs quite a lot of training data.Unfortunately,labeling these datasets will consume a lot of manpower,resources and time.As the results of these two challenges,it becomes more and more important to deal with how to solve scale-sensitive problem and how to conveniently prepare datasets.We delved into the challenges above.Aiming at the scale-sensitive problem of vehicle detection,we propose a scale-insensitive convolutional neural network(SINet)for fast vehicle detection.And we propose a domain adaptation algorithm for transferring vehicle detection model in different domains.Aiming at the conveniently preparing datasets,we build a new highway dataset,LSVH dataset.The main contents and contributions of this paper are as follows:1.We propose a scale-insensitive convolutional neural network(SINet).After research,we reveal two key issues: firstly,existing Ro I pooling layer in two-stage object detection network destroys the structure of small scale objects;On the other hand,the large intra-class distance for a large variance of scales exceeds the representation capability of a single network.Based on these two findings,we present a scale-insensitive convolutional neural network(SINet)for fast detecting multi-scale vehicles.1)we propose a context-aware Ro I pooling(CARoI pooling)layer to maintain the contextual information of small scale objects,which helps detecting small scale objects.2)we propose a multi-branch decision network to minimize the intra-class distance of features.The lightweight techniques bring zero extra time complexity but prominent detection accuracy improvement.The proposed techniques can be equipped with any deep network architectures of two-stage object detection and keep them trained end-to-end.2.Domain adaptation for preparing dataset conveniently.Because the process of video capture is unavoidably affected by various environmental factors,such as light intensity,weather(sunny or rainy),obstructive blur,and congestion.These factors can reflect the richness of the data set.Then,we propose a domain adaptation method of using one scene dataset that has been labeled,and transferring to another scene to obtain its annotation conveniently(such as sunny days to rainy days),thereby further reducing the workload of labeling datasets.3.We build a new highway dataset,LSVH dataset.LSVH not only has labeled data,but also has unlabeled data,which is beneficial for semi-supervised or unsupervised learning.As far as we know,our LSVH dataset is the first dataset for highway vehicle detection.LSVH dataset contains a large variance of scales vehicle objects,especially a lot of small vehicle objects.