In order to improve the accuracy of moving vehicle detection in intelligent transportation system, an accelerated proximal gradient-tensor recovery(APG-TR) algorithm was proposed based on tensor recovery. The traffic video image data were characterized by using tensor in the algorithm, which maintained the high-dimensional structure characteristic of video image. The lower rank part and sparse part in the tensor were effectively reconstructed by tensor recovery, and moving target vehicle and traffic background were separated, therefore the internal properties were easily extracted. The algorithm was tested by using 106 video images collected by traffic monitoring system. Test result shows that the average detection accuracies are 91.4% in fine days, 86.4% and 85.2% under rain and fog conditions respectively, which are more stable and accurate compared with the frame differential method. APG-TR algorithm is proved to have good convergence speed and robust, and has abroad application in the field of intelligent transportation.