With the steady progress in autonomous driving technology and the tremendous potential prospects for development, the topics about self-driving car have begun return to the center stage of AI applications. Hence, lots of efforts are made on algorithms relevant to self-driving car itself. However, few shed the light on the problem of how to testify them thoroughly, therefore unified standards for autonomous driving and testing is urgently needed. To study this problem, we begin by pedestrian detection, for that the ability of locating humans is one of the most critical problems that should be concerned about for self-driving cars. In this paper, we investigate to perform a standard evaluation to qualify different methods and detectors under a more practical manner. Specifically, we investigate several commonly used evaluation methodologies for pedestrian detection, and find out that the Caltech pedestrian detection benchmark is the most popular. However, it's still not suitable enough for the pedestrian detection problem in autonomous driving for not taking into account any emergency in practice. Besides, other benchmarks such as KITTI which evaluates the PASCAL-style mean Average Precision are also very inspirational. Our contributions are four-folds in this paper: (i) Summarizing existing pedestrian arts as thorough as possible, figure out the most convincing ones in the literature. (ii) Based on these most popular methods, we aim to propose a rational and unitary novel performance metrics tailor-made for evaluating the capacity of pedestrian detection for self-driving cars. (iii) Our testing standards take into account not only the general orientation of pedestrian detectors but also various practical application scenes. (iv) We also report the evaluation results of these promising pedestrian algorithms, to show where the gap still exists between the research work and the purpose of industrial product.