Joint rate adaptation and resource allocation for real-time H.265/HEVC video transmission over uplink OFDMA systems

Abstract

We consider multiuser video communication over uplink orthogonal frequency-division multiple access (OFDMA) systems. A cross-layer algorithm of joint bit allocation, packet scheduling and wireless resource assignment are proposed to minimize the end-to-end expected video distortion. Video rate adaptation is performed under the wireless resource constraints. The target number of encoding bits for each video packet is obtained to minimize the estimated distortion based on the online content-based rate-distortion function. Due to the inaccuracy of the rate control algorithm in H.265/HEVC encoding, the actual number of bits may differ from the target. Accordingly, the actual encoder distortion may deviate from the estimated distortion. Then, we propose an iterative algorithm to re-assign wireless resources based on the actual number of encoded bits to obtain the final resource allocation policy and packet scheduling decision. Numerical simulation results show that our proposed approach significantly outperforms the baseline algorithms in terms of received video quality.

Appendix: R-D parameter updating

We use the LMS method to derive the update formulation of the R-D parameters (4). For simplicity, we use D^est, a, b and r to represent $D_{k,m}^{est}$, $a_{k,m}^{est}$, $b_{k,m}^{est}$ and R_{k, m}/I_{k, m}, respectively. We take the logarithm for both side of (4).

$$ \begin{array}{@{}rcl@{}} &&{D^{est}} = a \cdot {r^{- b}}\\ &&\Rightarrow \ln {D^{est}} = \ln a - b \cdot \ln r\buildrel {\Delta} \over = a^{\prime} - b \cdot \ln r \end{array} $$

(27)

where a^′ = ln a. The squared error between the predicted distortion D^est and actual distortion D^act can be expressed as

$$ {e^{2}} = {\left( {\ln {D^{est}} - \ln {D^{act}}} \right)^{2}} $$

(28)

Taking the derivatives with respect to a^′ and b, we have

$$ \begin{array}{@{}rcl@{}} &&\frac{{\partial {e^{2}}}}{{\partial a^{\prime}}} = 2\left( {\ln {D^{est}} - \ln {D^{act}}} \right)\\ &&\frac{{\partial {e^{2}}}}{{\partial b}} = 2\left( {\ln {D^{est}} - \ln {D^{act}}} \right)\left( { - \ln r} \right) \end{array} $$

(29)

According to the adaptive Least Mean Square (LMS) method

$$ {a^{\prime}_{new}} = {a^{\prime}_{old}} - {\delta_{a}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right) $$

(30)

where δ_a is the update step of a. Therefore,

$$ \begin{array}{@{}rcl@{}} \ln {a_{new}} &=& \ln {a_{old}} - {\delta_{a}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right)\\ {a_{new}}& =& {a_{old}}{e^{- {\delta_{a}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right)}} \end{array} $$

(31)

After Taylor’s expansion and ignoring high-order terms,

$$ \begin{array}{@{}rcl@{}} {a_{new}} &=& {a_{old}}\left( {1 - {\delta_{a}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right)} \right)\\ &=& {a_{old}} - {\delta_{a}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right){a_{old}} \end{array} $$

(32)

For b,

$$ {b_{new}} = {b_{old}} + {\delta_{b}}\left( {\ln {D^{est}} - \ln {D^{act}}} \right)\ln r $$

(33)

where δ_b is the update step of b.