基于互相关的时序图像序列运动检测热力图异常排查

问题背景 #

我正在使用Python的互相关方法检测时序图像序列中的运动，通过地面固定相机的稳定视角识别突发运动（尤其是雪崩），最终将结果以热力图形式叠加在序列最后一帧图像上。

我已实现逐像素计算时间维度的互相关值，将结果存储为与原图像尺寸相同的矩阵，归一化后叠加生成热力图。但目前遇到异常：即使使用包含运动的序列，热力图仍显示单一颜色，结果不符合预期。

初始互相关计算代码 #

def compute_cross_correlation(image_sequence):
    frames, height, width = image_sequence.shape
    cross_corr_map = np.zeros((height, width))

    for i in range(height):
        for j in range(width):
            pixel_values = image_sequence[:, i, j]
            std = np.std(pixel_values)
            if std != 0:
                normalized_values = (pixel_values - np.mean(pixel_values)) / std
                cross_corr = np.correlate(normalized_values, normalized_values, mode='full')
                cross_corr_map[i, j] = np.max(np.abs(cross_corr))
            else:
                cross_corr_map[i, j] = np.nan

    return cross_corr_map

结果归一化代码 #

def normalize_cross_corr(cross_corr_map):
    min_val = np.min(cross_corr_map)
    max_val = np.max(cross_corr_map)
    normalized_map = (cross_corr_map - min_val) / (max_val - min_val)
    return normalized_map

更新后的互相关计算代码 #

基于建议修改了compute_cross_correlation函数，修改点：

• 获取互相关最大值时排除零延迟值
• 标准差为0时赋值为0

def compute_cross_correlation(image_sequence):
    frames, height, width = image_sequence.shape
    cross_corr_map = np.zeros((height, width))

    for i in range(height):
        for j in range(width):
            pixel_values = image_sequence[:, i, j]
            mean = np.mean(pixel_values)
            std = np.std(pixel_values)
            
            if std != 0:
                normalized_values = (pixel_values - mean) / std
                cross_corr = np.correlate(normalized_values, normalized_values, mode='full')
                
                max_cross_corr = np.max(cross_corr[len(cross_corr) // 2 + 1:])
                cross_corr_map[i, j] = max_cross_corr
            else:
                cross_corr_map[i, j] = 0  

    return cross_corr_map

内容的提问来源于stack exchange，提问作者joekim