This paper proposes a novel approach to estimate traffic flow velocity using data extracted from surveillance camera systems. This work deals with a practical issue where only a single camera is available and there are no pre-defined landmarks under the sense of that camera, thus we cannot identify the moving distance of a vehicle throughout time frames. To overcome this issue, we propose to utilize reference objects whose sizes are known in advance to estimate the pixel-per-meter ratio, then infer the moving distance of that vehicle within a time period which is used for its velocity estimation. We have conducted several empirical experiments using real-field datasets including images and videos from the traffic surveillance camera system of the Department of Transport of Ho Chi Minh City, Vietnam. The analysis results revealed that the proposed method was effective with the lowest bounding box and mask loss coefficient while the error rate was around 7.4% with our dataset and 7.47% with BrnoCompspeed. These results strongly suggest high applicability of the implementation in real world applications, especially in Vietnam.