We propose a method of objects tracking implemented on a smart image sensor. This method is able to detect not only the moving vector on a 2D plane but also the moving direction in a depth axis at very high speed such as more than 1000 fps. The proposed method uses not a normal block-based approach adopted by various smart sensors but a pixel-based approach using frame difference between neighbor pixels. Because of high inter-frame correlation at high frame rate, the 2D moving vector can be detected into a limited area which results in small column-parallel circuits on the sensor. In a depth axis, the moving direction is detected by object size changing. In this method, the detectable speed range can be controlled on the sensor and moved to lower speed by two methods. One is changing the reference timing for the frame difference from next frame to later frames and the other is changing the reference pixel position from neighbor pixels to interpolated pixels.