To measure the surface flow in a physical river model, an improved system of Large-Scale Particle Tracking Velocimetry (LSPTV) was proposed and the elements of the PTV system were described. Usually the tracer particles of a PTV system seeded on water surface tend to form conglomerates due to surface tension of water. In addition, they can not float on water surface when water flow is shallow. Ellipsoid particles were used to avoid the above problems. Another important issue is particle recognition. In order to eliminate the influence of noise, particles were recognized by the processing of multi-frame images. The kernel of the improved PTV system is the algorithm for particle tracking. A new 3-frame PTV algorithm was developed. The performance of this algorithm was compared with the conventional 4-frame PTV algorithm and 2-frame PTV algorithm by means of computer simulation using synthetically generated images. The results show that the new 3-frame PTV algorithm can recover more velocity vectors and have lower relative error. In addition, in order to attain the whole flow field from individual flow fields, the method of stitching individual flow fields by obvious marks was worked out. Then the improved PTV system was applied to the measurement of surface flow field in Model Yellow River and shows good performance.