python 聚类

python 聚类，

从小到大排序，返回小的完整索引。

def find_contiguous_indices(mask_all, max_gap=3):
    segments = []
    n = len(mask_all)
    i = 0
    while i < n:
        if mask_all[i]:
            start = i
            while i < n and mask_all[i]:
                i += 1
            end = i - 1
            segments.append((start, end))
        else:
            i += 1
    if not segments:
        return np.array([], dtype=int)  # 如果没有 True，返回空数组

    merged_segments = [segments[0]]
    for current in segments[1:]:
        last = merged_segments[-1]
        if current[0] - last[1] - 1 < max_gap:
            merged_segments[-1] = (last[0], current[1])  # 合并
        else:
            merged_segments.append(current)
    indices = []
    for start, end in merged_segments:
        indices.append([i for i in range(start, end + 1)])  # 添加该段所有索引

    return indices
	

            best_k = 5  # 通过肘部法则或轮廓系数确定，此处简化为已知值

            # 执行KMeans++聚类
            kmeans = KMeans(n_clusters=best_k, init='k-means++', n_init=10)
            labels = kmeans.fit_predict(speeds.reshape(-1, 1))

            # unique_labels = np.unique(labels)
            centers_flattened=kmeans.cluster_centers_.flatten()
            # 按聚类中心从大到小排序
            sorted_indices = np.argsort(centers_flattened)

            mask_all = np.zeros_like(labels, dtype=bool)

            for idx in sorted_indices:
                mask = (labels == idx)
                mask_all |= mask  # 等价于 mask_all = mask_all | mask
                if np.sum(mask_all == True)>len(pose_3d)*0.15 and centers_flattened[idx]>1600:
                    break
            low_frame_ids=find_contiguous_indices(mask_all)