Aiming at the problems of the traditional video traffic flow detection algorithm, such as single feature, poor weather robustness and easy to be blocked by vehicles, a traffic flow detection method combining temporal and spatial features was proposed. To the cross-correlation of foreground and background image blocking normalized values as a feature of testing vehicles through accumulation plan get space-time characteristic figure on time, on time accumulated figure to analysis the characteristic value, and the introduction of deep learning target detection algorithm assisted background updating and eliminate the problem such as vehicles for shade caused by the leak, and correction for pedestrians, bicycles and other interference factors to achieve the objective of the statistics of the number of cars. Experimental results show that,the accuracy of the algorithm in this paper can generally reach more than 91%, which is better than the traditional frame difference method and optical flow method in most cases. Moreover, the real-time performance is good, which can better meet the requirements of the traffic flow detection system.