Pedestrian tracking in vehicle coordinates is not only a necessary part of the perception for autonomous cars but also a challenging task in computer vision. This paper presents a multi-sensor fusion model combining images and laser scanning data to track pedestrians in occupied grid maps. In our approach, the bounding boxes of pedestrians detected in images are used to generate regions of interest (ROI) in grids. A data association method based on multi-characteristic Mahalanobis distance (MMD) and sliding windows is proposed to establish correspondence between the detections in different frames. Then Sampling Importance Resampling Particle Filter (SIR PF) is used to update pedestrians' states with the fusion of images and grids. Experiments show our method has competitive efficiency and accuracy compared to conventional methods. Codes of this paper is released.