2.2 Motion Detection
Before the features are calculated, a motion-detection process is applied to the input sequence, as shown in Fig. 1,to check whether the targets are in motion. Motion detection is performed by first generating a frame-difference image between the input image and its neighboring image. Each pixel value in the difference image and is tested against a predefined threshold to identify the pixel locations at which the intensity changes abruptly. Using the motion-detection process, we can obtain information on the size and speed of the target regions in motion. If the targets are in slow
motion and relatively small and the backgrounds are stationary, a frame-difference image is generated to better
capture the moving targets. Otherwise, a frame image is used.


2.3 Feature Extraction
Each of the features is calculated within a local region, using the dual rectangular window for every pixel in the image.An approximate target size is predetermined via prior knowledge of the range and filed of view (FOV) information. The size of the inner and outer windows is set according to the approximate target size. We also assume the length
of the longest vehicle and the height of the tallest vehicle in the target set to be known. The first two features (the local maximum and the local contrast measure) are suitable to detect high-contrast targets, while the rest (the local average gradient strength and the local variation) are designed to find internal intensity variations.

2.2 动作探测
在考虑特性之前, 一种动作探测程序适合于输入系列，如图1所示，检查目标是否在运动中。动作探测表现为首先在输入的图像与其参照物之间产生一种不同画面的图像。各像素评估不同的图像，并根据预先设定的极限来进行测试，据此识别图像区域中的强度的骤变。使用动作探测程序，我们可以得到动作目标区域在大小和速度方面的信息。如果目标在动作方面慢，并且相对少，背景也就固定，就能较好地捕捉到运动中的目标而生成(背景)不同画面的图像。反之，则需要使用一种结构成像。
2.3特性筛选
使用双重矩形窗，图像中的全部象素的各种特性被设计在一个局部区域内。通过对图像已知的范围和设定的取景数据（FOV），预先确定了一个近似的目标值。内视窗与外视窗的内外尺寸被按照近似目标值设定。我们还通过设想车辆最长的长度和最高的高度来设定已知目标。最先的两个特性（局部最大值和局部对比度）都适合于探测对比度强烈的目标。而其它（局部平均倾斜的强度和局部变化）则被设计为寻找内在强度变化。