Fast and precise measurements of live single-cell biophysical properties is significant in disease diagnosis, cytopathologic analysis, etc. Existing methods still suffer from unsatisfied measurement accuracy and low efficiency. We propose a computer vision method to track cell dielectrophoretic movements on a microchip, enabling efficient and accurate measurement of biophysical parameters of live single cells, including cell radius, cytoplasm conductivity, and cell-specific membrane capacitance, and in situ extraction of cell texture features. We propose a prediction-iteration method to optimize the cell parameter measurement, achieving high accuracy (less than 0.79% error) and high efficiency (less than 30 s). We further propose a hierarchical classifier based on a support vector machine and implement cell classification using acquired cell physical parameters and texture features, achieving high classification accuracies for identifying cell lines from different tissues, tumor and normal cells, different tumor cells, different leukemia cells, and tumor cells with different malignancies. The method is label-free and biocompatible, allowing further live cell studies on a chip, e.g., cell therapy, cell differentiation, etc.