手动解析 H.264 SPS/PPS 参数集

手动解析 H.264 SPS/PPS 参数集

引言

在处理 H.264 视频流时，SPS（序列参数集）和 PPS（图像参数集）包含了解码视频所需的关键信息。

当我们在 WebRTC 等场景中遇到解码的疑难问题时，手动解析这些参数集往往能帮助我们找出问题根源。

本文不是 H.264 流的深入教学，而是记录一下手动解析 SPS/PPS 的过程，解析过程可能会有些枯燥，但是完成一遍之后，可以举一反三，方便大家对照和学习解析机制。

如果你有需要，nalu.qer.im 是我做的在线解析 .h264 文件的网站，也依托了这些解析原理，欢迎使用。

必备知识

在开始解析前，需要简单了解以下基础概念：

（1）H.264 流格式基础：H.264 流由 NALU（Network Abstraction Layer Unit）组成。NAL U 是不定长的，每个 NALU 前有起始码，标准起始码为 0x000001 或 0x00000001，这样从每个起始码开始，直到下一个起始码或者结尾之前，则为一个 NALU。

（2）NALU 结构：去掉起始码后，NALU 本身由 NAL Header 和 NAL Payload 两部分组成。NAL Header 中指明了当前 NALU 的类型，本文将重点解析 SPS 和 PPS 类型的 NALU。

（3）防竞争字节：标准起始码为 0x000001 或 0x00000001，但是如果 NALU 数据部分（Payload）中恰好也含有相同的序列，解码器可能错误地将其识别为起始码，导致解析错误。

编码时：编码时会先扫描一遍数据部分，当遇到两个连续的 0x00 字节（即 0x00 0x00），且第三个字节为 0x00、0x01、0x02 或 0x03 时，会在第三个字节前插入 0x03，使其变为 0x00 0x00 0x03 [原第三个字节]。如把 0x00 0x00 0x01 变为 0x00 0x00 0x03 0x01。

解码时：扫描数据部分，遇到 0x00 0x00 0x03 时，移除第三个字节（0x03）。如把 0x00 0x00 0x03 0x01 变为 0x00 0x00 0x01。

（4）按字节顺序读取：获取处理后的字节流后，我们按照H.264标准规定的顺序解析各参数。标准中对每个参数都有明确的位长度标注：

  • 固定长度参数：如 u(1) 表示读取1位，u(8) 表示读取 8 位（1字节），读取后转换为十进制值
  • 可变长度参数：标注为 ue(v) 或 se(v) 的参数需要按照指数哥伦布编码方式解析

指数哥伦布编码是H.264中广泛采用的一种高效变长编码方式，相比固定位数编码，它在表示大数值时能显著节省空间。主要分为两种类型：

无符号指数哥伦布编码 ue(v)：用于编码非负整数

  • 读取比特流中连续的0，直到遇到第一个1为止，计数0的个数为M
  • 接着读取后续M个比特，转换为十进制数N
  • 最终解码值 = 2^M - 1 + N

有符号指数哥伦布编码 se(v)：用于编码带符号整数

  • 先按 ue(v) 方法解码得到中间值 k
  • 根据k的奇偶性确定最终值：
    • k 为奇数：se(v) = (k+1)/2（正数）
    • k 为偶数：se(v) = -k/2（负数）

解析实例 1：SPS 类型的 NALU

预处理：原始数据

下面是一个原始的 NALU，以十六进制展示为：

00 00 00 01 67 64 00 28 ac d9 40 78 02 27 e5 c0 44 00 00 03 00 04 00 00 03 00 e8 3c 60 c6 58

去掉防竞争字节（0x00 0x00 0x03 -> 0x00 0x00）后：

00 00 00 01 67 64 00 28 ac d9 40 78 02 27 e5 c0 44 00 00 00 04 00 00 00 e8 3c 60 c6 58

转换为比特流后：(转换工具)

0000000000000000000000000000000101100111011001000000000000101000101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000

解析：NAL 分隔符

分隔符 -> [00000000000000000000000000000001]01100111011001000000000000101000101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000

解析：NAL Header

NAL Header 一共 8 位，每一位的含义如下：

+---------------+
|0|1|2|3|4|5|6|7|
+-+-+-+-+-+-+-+-+
|F|NRI|  Type   |
+---------------+

其中

F 为 forbidden_zero_bit：禁止位，初始为0，当网络发现NAL单元有比特错误时可设置该比特为1，以便接收方纠错或丢掉该单元。

NRI 为 nal_ref_idc：NAL 的重要性指示，标志该 NAL 单元的重要性，值越大，越重要，解码器在解码处理不过来的时候，可以丢掉重要性为 0 的 NALU。

Type 为 nal_unit_type：转换为十进制之后，不同的值代表不同的含义，下面是对应的表格

NAL单元类型表

取 8 位： NAL Header -> [01100111]011001000000000000101000101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000

解析这 8 位： 0110 0111 0… … # forbidden_zero_bit -> 0 -> u(1) .11. … # nal_ref_idc —>u(2) -> 2 -> HIGHEST …0 0111 # nal_unit_type -> u(5) -> 7 -> SPS

解析：NAL Payload

（为了防止出错，所有的 bit 流我都完整的书写了，请理解）

H.264 标准协议中规定的 SPS 格式位于文档的 7.3.2.1.1 部分，以下解析按照此表进行

SPS参数集格式

profile_idc -> u(8) -> [01100100]0000000000101000101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 100

constraint_set_flag + reserved_zero_2bits -> u(8) -> [00000000]00101000101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 0000 0000 0… … #constraint_set0_flag .0.. … #constraint_set1_flag ..0. … #constraint_set2_flag …0 … #constraint_set3_flag … 0… #constraint_set4_flag … .0.. #constraint_set5_flag … ..00 #reserved_zero_2bits

level_idc -> u(8) -> [00101000]101011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 40(Supports 2Kx1K format. Enables Interlace support. 62914560 samples/sec)

seq_parameter_set_id -> ue(v) -> 1[]01011001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0

if (profile_idc == 100) -> true chroma_format_idc -> ue(v) -> 01[0]11001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=1,N=0 -> ans=2^M-1+N=1 if (chroma_format_idc == 3) -> false

bit_depth_luma_minus8 -> ue(v) -> 1[]1001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0

bit_depth_chroma_minus8 -> ue(v) -> 1[]001101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0

qpprime_y_zero_transform_bypass_flag -> ue(1) -> [0]01101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0

seq_scaling_matrix_present_flag -> ue(1) -> [0]1101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0

if (seq_scaling_matrix_present_flag) -> false log2_max_frame_num_minus4 -> ue(v) -> 1[]101100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0

pic_order_cnt_type -> ue(v) -> 1[]01100101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0

if (pic_order_cnt_type == 0) -> true log2_max_pic_order_cnt_lsb_minus4 -> ue(v) -> 01[1]00101000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=1,N=1 -> ans=2^M-1+N=2

max_num_ref_frames -> ue(v) -> 001[01]000000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=2,N=1 -> ans=2^M-1+N=4

gaps_in_frame_num_value_allowed_flag -> u(1) -> [0]00000011110000000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0

pic_width_in_mbs_minus1 -> ue(v) -> 0000001[111000]0000001000100111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=6,N=56(111000) -> ans=2^M-1+N=119 -> frame_with = 16*pic_width_in_mbs = 16*(pic_width_in_mbs_minus1+1) = 16*(119+1)=1920

pic_height_in_map_units_minus1 -> ue(v) -> 0000001[000100]111111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=6,N=4(000100) -> ans=2^M-1+N=67 -> frame_height = 16*68 = 1088

frame_mbs_only_flag -> u(1) -> [1]11111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1

if (!frame_mbs_only_flag) -> false

direct_8x8_inference_flag -> u(1) -> [1]1111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1

frame_cropping_flag -> u(1) -> [1]111001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1

if (frame_cropping_flag) -> true frame_crop_left_offset -> ue(v) -> 1[]11001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0 frame_crop_right_offset -> ue(v) -> 1[]1001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0 frame_crop_top_offset -> ue(v) -> 1[]001011100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0 frame_crop_bottom_offset -> ue(v) -> 001[01]1100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> M=2,N=1(01) -> ans=2^M-1+N=4

vui_parameters_present_flag -> u(1) -> [1]100000001000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1

if (vui_parameters_present_flag) -> true vui_parameters() // E.1.1 VUI parameters syntax Page.461 aspect_ratio_info_present_flag -> u(1) -> [1]000000010001 00000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1 if (aspect_ratio_info_present_flag) -> true aspect_ratio_idc -> u(8) -> [00000001]000100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1 if (aspect_ratio_idc == 255) -> false overscan_info_present_flag -> u(1) -> [0]00100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0 if (overscan_info_present_flag) -> false video_signal_type_present_flag -> u(1) -> [0]0100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0 if (video_signal_type_present_flag) -> false chroma_loc_info_present_flag -> u(1) -> [0]100000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 0 timing_info_present_flag -> u(1) -> [1]00000000000000000000000000000001000000000000000000000000001110100000111100011000001100011001011000 -> 1 if (timing_info_present_flag) -> true num_units_in_tick -> u(32) -> [00000000000000000000000000000001]000000000000000000000000001110100000111100011000001100011001011000 -> 1 time_scale -> u(32) -> u(32) -> [00000000000000000000000000111010]0000111100011000001100011001011000 -> 58 fixed_frame_rate_flag -> u(1) -> [0]000111100011000001100011001011000 -> 0 nal_hrd_parameters_present_flag -> u(1) -> [0]00111100011000001100011001011000 -> 0 if (nal_hrd_parameters_present_flag) -> false vcl_hrd_parameters_present_flag -> u(1) -> [0]0111100011000001100011001011000 -> 0 if (vcl_hrd_parameters_present_flag) -> false if(nal_hrd_parameters_present_flag || vcl_hrd_parameters_present_flag) -> false pic_struct_present_flag -> u(1) -> [0]111100011000001100011001011000 -> 0 bitstream_restriction_flag -> u(1) -> [1]11100011000001100011001011000 -> 1 if(bitstream_restriction_flag ) -> true motion_vectors_over_pic_boundaries_flag -> u(1) -> [1]1100011000001100011001011000 -> 1 max_bytes_per_pic_denom -> ue(v) -> 1[]100011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0 max_bits_per_mb_denom -> ue(v) -> 1[]00011000001100011001011000 -> M=0,N=0 -> ans=2^M-1+N=0 log2_max_mv_length_horizontal -> ue(v) -> 0001[100]0001100011001011000 -> M=3,N=4 -> ans=2^M-1+N=11 log2_max_mv_length_vertical -> ue(v) -> 0001[100]011001011000 -> M=3,N=4 -> ans=2^M-1+N=11 max_num_reorder_frames -> ue(v) -> 01[1]001011000 -> M=1,N=1 -> ans=2^M-1+N=2 max_dec_frame_buffering -> ue(v) -> 001[01]1000 -> M=2,N=1 -> ans=2^M-1+N=4 rbsp_trailing_bits() // 推断 rbsp_stop_one_bit -> u(1) -> [1]000 -> 1 rbsp_alignment_zero_bit -> u(1) -> [0]00 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0]0 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0] -> 0

解析实例 2：PPS 类型的 NALU

预处理：原始数据

下面是一个原始的 NALU，以十六进制展示为：

00 00 00 01 68 eb e3 cb 22 c0

无防竞争字节。

转换为比特流后：

00000000000000000000000000000001011010001110101111100011110010110010001011000000

解析：NAL 分隔符

分隔符 -> [00000000000000000000000000000001]011010001110101111100011110010110010001011000000

解析：NAL Header

取 8 位： NAL Header -> [01101000]1110101111100011110010110010001011000000

0110 1000 0… … # forbidden_zero_bit -> u(1) .11. … # nal_ref_idc -> u(2) -> HIGHEST …0 1000 # nal_unit_type -> u(5) -> PPS

解析：NAL Payload

在 H.264 的协议文档中，PPS 的结构定义在 7.3.2.2 节中，具体的结构如下表所示：

PPS参数集格式

pic_parameter_set_id -> ue(v) -> 1[]110101111100011110010110010001011000000 -> 0 seq_parameter_set_id -> ue(v) -> 1[]10101111100011110010110010001011000000 -> 0

entropy_coding_mode_flag -> [1]0101111100011110010110010001011000000 -> 1

bottom_field_pic_order_in_frame_present_flag -> [0]101111100011110010110010001011000000 -> 0

num_slice_groups_minus1 -> ue(v) -> 1[]01111100011110010110010001011000000 -> 0

if (num_slice_groups_minus1 > 0) -> false

num_ref_idx_l0_default_active_minus1 -> ue(v) -> 01[1]11100011110010110010001011000000 -> M=1,N=1 -> ans=2^M-1+N=2

num_ref_idx_l1_default_active_minus1 -> ue(v) -> 1[]1100011110010110010001011000000 -> M=0,N=0 -> ans=2^M-1+N=0

weighted_pred_flag -> u(1) -> [1]100011110010110010001011000000 -> 1

weighted_bipred_idc -> u(2) -> [10]0011110010110010001011000000 -> 2

pic_init_qp_minus26 -> se(v) -> 001[11]10010110010001011000000 -> M=2,N=3 -> code_num=2^M-1+N=6 -> (偶数)ans=-code_num / 2 = -3

pic_init_qs_minus26 -> se(v) -> 1[]0010110010001011000000 -> M=0,N=0 -> code_num=2^M-1+N=0 -> (偶数)ans=-code_num / 2 = 0

chroma_qp_index_offset -> se(v) -> 001[01]10010001011000000 -> M=2,N=1 -> code_num=2^M-1+N=4 -> (偶数)ans=-code_num / 2 = -2

deblocking_filter_control_present_flag -> u(1) -> [1]0010001011000000 -> 1

constrained_intra_pred_flag -> u(1) -> [0]010001011000000 -> 0

redundant_pic_cnt_present_flag -> u(1) -> [0]10001011000000 -> 0

transform_8x8_mode_flag -> u(1) -> [1]0001011000000 -> 1

pic_scaling_matrix_present_flag -> u(1) -> [0]001011000000 -> 0

if(pic_scaling_matrix_present_flag) -> false

second_chroma_qp_index_offset -> se(v) -> 001[01]1000000 -> M=2,N=1 -> code_num=2^M-1+N=4 -> (偶数)ans=-code_num / 2 = -2

rbsp_trailing_bits() rbsp_stop_one_bit -> u(1) -> [1]000000 -> 1 rbsp_alignment_zero_bit -> u(1) -> [0]00000 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0]0000 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0]000 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0]00 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0]0 -> 0 rbsp_alignment_zero_bit -> u(1) -> [0] -> 0

小结

终于解析完了上述两个实例，十分舒爽，从此遇到任何解析都不怕了（希望是）。感谢阅读，再补充一些解码视频所必需的关键信息：

SPS中的关键参数及其意义

  1. 分辨率相关参数：

    • pic_width_in_mbs_minus1：确定视频宽度，实际像素宽度 = (pic_width_in_mbs_minus1 + 1) × 16
    • pic_height_in_map_units_minus1：确定视频高度，实际像素高度 = (pic_height_in_map_units_minus1 + 1) × 16
    • frame_cropping_flag和相关裁剪参数：用于微调最终显示的视频区域

  2. 视频特性参数：

    • profile_idc：指示H.264编码配置文件（如Baseline、Main、High等）
    • level_idc：指示视频复杂度级别，影响最大分辨率、帧率和码率
    • chroma_format_idc：指示色彩空间格式（如YUV 4:2:0、4:2:2等）

  3. 时序参数：

    • timing_info_present_flag：是否包含时序信息
    • num_units_in_tick和time_scale：用于计算视频帧率，帧率 = time_scale / (2 × num_units_in_tick)

PPS中的关键参数及其意义

  1. 编码控制参数：

    • entropy_coding_mode_flag：熵编码模式（0为CAVLC，1为CABAC）
    • pic_init_qp_minus26：初始量化参数，影响视频质量和压缩率
    • deblocking_filter_control_present_flag：是否启用去块滤波控制

  2. 参考帧参数：

    • num_ref_idx_l0/l1_default_active_minus1：默认活动参考帧索引数量
    • weighted_pred_flag和weighted_bipred_idc：加权预测相关标志

（全文完）