Since performance of traditional machine learning methods of detecting vehicles in traffic surveillance video is influenced by objective factors such as video quality, shooting angle and weather, which results in complex preprocessing, hard generalization and poor robustness, combined with dilated convolution, feature pyramid and focal loss, two deep learning models which are improved Faster R-CNN (Faster Regions with Convolutional Neural Network) and SSD (Single Shot multibox Detector) model were proposed for vehicle detection. Firstly, a dataset was composed of 851 labeled images captured from the surveillance video at different time. Secondly, improved and original models were trained under same training strategies. Finally, average accuracy of each model were calculated to evaluate. Experimental results show that compared with original Faster R-CNN and SSD, the average accuracies of the improved models improve 0.8 percentage points and 1.7 percentage points respectively. Both deep learning methods are more suitable for vehicle detection in complicated situation than traditional methods. The former has higher accuracy and slower speed, which is more suitable for video off-line processing, while the latter has lower accuracy and higher speed, which is more suitable for video real-time detection.