PaperlayerView exact revision on arXiv

Source-linked AI summary

CRB-Guided Sensing and Resource Allocation for Human Pose Prediction in Integrated Sensing, Communication, and Computation Systems

Zhonghao Liu, Yahao Ding, Jiaxiang Wang, Zhaohui Yang, Abdol Hamid Aghvami, Mohammad Shikh-Bahaei

arXiv:2609.01908v1cs.ITcs.NI

TL;DR

Resource-constrained ISCC systems lack a quantitative link between sensing uncertainty, limited inference depth, and human pose prediction error. The paper develops a CRB-guided ET-Mamba framework with adaptive-depth prediction and joint resource optimization, reducing MPJPE by up to approximately 35% compared with the baseline under resource constraints.

  • Problem

    Existing studies lack quantitative analysis linking point-cloud perturbation, computation-limited inference, and sensing-task performance in resource-constrained mmWave ISCC systems.

  • Method

    The paper combines ET-Mamba with adaptive-depth prediction, CRB-guided point-cloud perturbations, an empirical SNR-depth-error model, and AO-based joint optimization of beamforming, inference depth, and computation frequency.

  • Results

    The proposed optimization reduces MPJPE by up to approximately 35% compared with the baseline under resource constraints.

  • Takeaways & Limitations

    Joint sensing, communication, and computation design improves pose prediction performance in resource-constrained ISCC systems.

Abstract

from arXiv · show

Integrated sensing, communication, and computation (ISCC) provides a promising framework for indoor human-centric applications. In these applications, short-term human pose prediction facilitates continuous human pose tracking and proactive resource allocation. This paper proposes a Cramer-Rao bound (CRB)-guided sensing framework and investigates a problem of minimizing prediction error in resource-constrained ISCC systems. Specifically, a pose prediction model (ET-Mamba) is first developed to predict human joint positions for continuous tracking. To account for computation-resource limitations, lightweight prediction heads are attached to different inference layers, enabling adaptive-depth pose prediction. A CRB-guided perturbation strategy is then introduced to translate sensing uncertainty at different sensing SNR levels into point-cloud perturbations. Based on that, an empirical relationship among pose prediction error, sensing SNR, and model inference depth is established. Furthermore, to improve prediction accuracy under limited resources, this paper formulates a resource allocation optimization problem that minimizes the pose prediction error by jointly optimizing the beamforming matrix, model inference depth, and computation frequency. To solve this mixed-integer non-convex optimization problem, we propose an alternating optimization (AO)-based algorithm, where closed-form updates and semidefinite programming (SDP) are integrated into the iterative solution process. Simulation results show that the proposed method effectively improves pose prediction performance by up to 35 percent under resource constraints, verifying the effectiveness of conducting joint sensing, communication, and computation design in ISCC systems.

I. Introduction

The introduction identifies a gap in linking sensing uncertainty and computation-limited inference to pose prediction error, then proposes a CRB-guided framework with joint resource allocation for indoor mmWave ISCC systems.

  • Existing physical-layer studies optimize communication or sensing metrics, while learning-based methods largely overlook how limited resources affect downstream pose prediction error.
  • Resource allocation can induce sensing-SNR fluctuations and point-cloud jittering, while computation limits may prevent full-depth inference within an operating slot.
  • The proposed ET-Mamba predicts future joint positions from historical mmWave point-cloud sequences using lightweight heads for adaptive model-depth execution.
  • The framework formulates joint optimization of point-cloud perturbation, model inference depth, and resource allocation to minimize total pose prediction error.
  • A sigmoidvariant empirical model links sensing SNR, inference depth, and pose error, while an AO algorithm combines closed-form solutions with SDP.
  • The proposed optimization reduces MPJPE by up to approximately 35% compared with the baseline under resource constraints.

A. Communication model

The communication model represents an indoor mmWave ISCC access point serving communication users while transmitting sensing signals, with propagation modeled through array responses, path loss, and Rician fading.

  • The UPA uses standard half-wavelength element spacing, d = λ/2, with d and λ denoting element spacing and carrier wavelength.
  • The communication signal combines user data and a sensing waveform through communication and sensing beamforming vectors.
  • The received user signal includes the downlink channel and Gaussian noise, with the channel accounting for both line-of-sight and non-line-of-sight propagation.
  • The achievable communication rate is defined as Ru = Bc log2(1 + SINRu).
  • The sensing model treats the human body as an extended target whose visible surface is partitioned into non-overlapping scattering regions.
  • Static clutter is suppressed before CFAR detection and FFT processing extract range, Doppler, and angle information for point-cloud construction.

III. CRB-Guided Pose Prediction

The CRB-guided pose prediction framework combines ET-Mamba with geometry-aware point-cloud processing and adaptive-depth inference. It preserves spatial and temporal structure while supporting flexible computation-accuracy trade-offs.

  • ET-Mamba architecture: ET-Mamba uses Mamba sequence modeling while preserving geometric relationships in unordered, sparse mmWave point clouds.Geometry distance ordering is introduced before sequence modeling to reduce the weakening of local spatial relationships caused by naive serialization.
  • ET-Mamba architecture: Global-local fusion adaptively injects whole-frame context into local group tokens while retaining neighborhood-level geometry.A learned gate controls the contribution of global context before residual refinement of the fused token.
  • Geometry-aware serialization: Two reference-point orderings encode complementary spatial structure relative to the radar origin and the centroid of group centers.The resulting ordered sequences are concatenated, producing 2G tokens per frame.
  • Geometry-aware serialization: Interleaving tokens across frames places temporally related groups closer together, helping Mamba capture cross-frame motion correlations.The serialized sequence is formed by interleaving corresponding ordered group tokens from different time steps.
  • Adaptive-depth inference: Lightweight prediction heads after each of six Mamba layers enable candidate inference depths when full-depth execution is infeasible.Adaptive-depth execution provides a flexible trade-off between inference accuracy and server-side power or latency constraints.

B. CRB-Guided Perturbation and Training

The perturbation strategy models sensing uncertainty from resource-dependent SNR using CRB-based noise in range, angle, and radial velocity. Augmented sequences train the predictor to handle varying sensing quality.

  • CRB-guided perturbation: CRB-guided anisotropic perturbations emulate point-cloud degradation caused by changing sensing power and beamforming conditions.Range, angle, and radial-velocity perturbations are modeled as independent Gaussian variables with CRB-derived variances.
  • SNR-dependent uncertainty: Per-point SNR is obtained from beam-center SNR using beamforming gain and distance-dependent path-loss correction.The correction uses the point range and the distance from the radar to the beam center.
  • Point-cloud augmentation: Perturbed range and angular parameters are transformed back into Cartesian coordinates, and radial-velocity noise augments each point feature.The spherical-to-Cartesian mapping produces perturbed coordinates before the augmented feature is formed.
  • Training procedure: Sliding-window samples contain M consecutive point-cloud frames and predict the 3D joint positions in the next frame.The network is optimized with Adam using mean squared error over predicted and ground-truth joint positions.

IV. PERFORMANCE ANALYSIS AND PROBLEM FORMULATION

The system jointly models communication, sensing, and computation requirements through QoS, latency, and energy constraints. Its resource formulation accounts for beamforming, point-cloud generation, and inference execution.

  • IV. PERFORMANCE ANALYSIS AND PROBLEM FORMULATION: The optimization jointly selects beamforming, model depth, and computation frequency while minimizing MPJPE for continuous pose tracking.The formulation adapts sensing, communication, and computation resources to communication QoS and sensing-performance requirements.
  • A. Energy Consumption and Latency: ISCC energy consumption comprises sensing transmission, communication transmission, and local computation energy.Local computation includes point-cloud generation and model inference.
  • A. Energy Consumption and Latency: Point-cloud generation computes FFTs across sensing snapshots, with energy and latency determined by computational intensity and allocated CPU frequency.The point-cloud generation model uses an Ns-point FFT for each of Nc sensing snapshots.
  • A. Energy Consumption and Latency: Inference energy and latency depend on depth-independent processing, the final prediction head, and the CPU cycles required by each depth layer.The formulation distinguishes global processing, group construction, prediction-head, and per-layer computational costs.
  • B. Communication and sensing QoS: The communication users are quasi-static within each slot, with locations updated between slots, and each user must satisfy Ru ≥ Rmin.An initial rapid scan estimates the sensing target location for subsequent beamforming and tracking.
  • B. Communication and sensing QoS: Sensing QoS requires coverage of all human joints, reliable detection of the body center, and balanced illumination across visible scatterers.The detection requirement is expressed through a minimum sensing SNR threshold at the human center.

C. Empirical Model of MPJPE

The paper fits an empirical MPJPE model that captures the effects of sensing SNR, inference depth, and their coupling. This model serves as the objective for joint resource allocation solved by alternating optimization.

  • Empirical relationship: MPJPE generally decreases with both sensing SNR and model inference depth.Lower SNR produces less reliable point-cloud features and larger prediction error, while deeper inference usually strengthens feature extraction.
  • Empirical relationship: A sigmoid-based model captures nonlinear, saturating effects of SNR and inference depth on MPJPE.The fitted relationship includes constant, SNR, inference-depth, and SNR-depth coupling terms.
  • Problem formulation: The fitted MPJPE function is used as the final system-performance representation in the pose prediction error minimization problem.The optimization objective depends on sensing SNR and selected model inference depth.
  • Problem formulation: The optimization enforces time-slot completion, energy budget, communication QoS, minimum sensing SNR, depth bounds, and computation-frequency bounds.These constraints jointly limit feasible sensing, communication, and computation decisions.
  • Alternating optimization: Alternating optimization decomposes the mixed-integer non-convex problem into computation-resource and beamforming subproblems.The two subproblems alternately optimize computation variables and beamforming variables, yielding a stationary solution.

1) Computation Resource Allocation With Fixed Beamforming Matrix:

With fixed beamforming, computation allocation exploits monotonicity to maximize feasible inference depth and CPU frequencies while satisfying latency, energy, and depth constraints.

  • Depth selection: The sensing error m(C) decreases with inference depth, so the optimal depth is the largest feasible integer C.The continuous relaxation is projected back onto the feasible integer set after optimization.
  • Optimality conditions: The latency constraint is active at the optimum because reducing feasible CPU frequencies lowers energy and permits greater inference depth.This follows from the computation energy increasing monotonically with fpc and fin.
  • Point-cloud generation frequency: For fixed fin, fpc is maximized because both inference depth and prediction error vary monotonically with fpc.The resulting optimum lies at the boundary imposed by the feasible constraints.
  • Inference frequency: For fixed fpc, fin is likewise selected as the maximum feasible value satisfying all constraints simultaneously.Its upper bounds arise from the energy and model-depth constraints.
  • Discrete recovery: After continuous depth optimization, the algorithm selects the largest feasible integer depth and re-optimizes CPU frequencies to minimize computation energy.Any remaining budget can then be reallocated to beamforming.

2) Beamforming Matrix Design With Fixed Computation Resources:

With computation resources fixed, beamforming is designed through semidefinite relaxation so sensing performance improves while communication QoS constraints remain satisfied.

  • Problem formulation: The beamforming subproblem is formulated after fixing computation frequencies and inference depth.The remaining computation energy determines the budget available to beamforming.
  • Semidefinite relaxation: Semidefinite relaxation replaces beamforming vectors with positive semidefinite matrix variables Wc,u and Wr.Rank-one beamformers are recovered afterward using Gaussian randomization.
  • Sensing objective: Sensing SNR is expressed through κs Tr(WrAo), linking the sensing beamforming matrix to the sensing objective.The matrix Ao is formed from the sensing steering vector.
  • Sensing objective: Because the fitted error decreases with Tr(WrAo), optimal beamforming maximizes its feasible value.The fitted error function is monotonically decreasing in this scalar quantity.
  • QoS-constrained design: Communication QoS becomes fixed SINR threshold constraints, yielding an SDP that can be solved efficiently with CVX.The AO algorithm alternates this beamforming step with computation-resource updates.

B. Algorithm Analysis

The AO algorithm alternates computation-resource and beamforming updates, with each step preserving feasibility and not increasing the fitted pose-prediction error.

  • Convergence behavior: Each AO iteration updates CPU frequencies and model depth, then solves an SDP beamforming problem to improve sensing SNR.The computation and beamforming blocks are optimized within their feasible regions.
  • Convergence behavior: Monotonic decreases in fitted error with model depth and sensing SNR ensure that AO updates do not increase the objective value.The non-negativity of MPJPE supports the resulting convergence argument.

1) Complexity Analysis:

The proposed AO method has lower computational complexity than exhaustive search over feasible model inference depths.

  • Comparison: Exhaustive search checks candidate inference depths and solves beamforming only for feasible candidates, producing higher complexity than the AO algorithm.NC denotes the number of feasible depths among Cmax candidates.

VI. Results and Analysis

The evaluation uses MM-Fi sequences and MPJPE-based metrics to compare ET-Mamba with representative point-cloud learning baselines. ET-Mamba achieves the strongest reported pose-prediction accuracy while maintaining favorable computational complexity.

  • Experimental Settings: MM-Fi sequences use eight-frame point-cloud histories to predict 3D human joint positions in the next frame.The dataset provides mmWave radar point clouds and depth-camera pose annotations.
  • Evaluation Metrics: MPJPE and PA-MPJPE measure predicted-joint accuracy, with lower values indicating better pose prediction.PA-MPJPE first applies Procrustes alignment to reduce global misalignment effects.
  • Quantitative Comparison: ET-Mamba reduces MPJPE from 3.85 cm to 3.24 cm and PA-MPJPE from 2.42 cm to 2.20 cm versus STPM.These comparisons use the same experimental protocol as the representative baselines.
  • Quantitative Comparison: ET-Mamba achieves the best MPJPE and PA-MPJPE among the compared methods while balancing prediction accuracy and computational complexity.The comparison includes model parameters and FLOPs in addition to pose-error metrics.

1) Quantitative Comparison:

The results evaluate robustness and resource allocation across sensing quality, power, sensing duration, communication rate, inference frequency, and model depth. Adaptive joint allocation and CRB-guided perturbation consistently improve pose prediction under constrained conditions, with gains reaching approximately 35%.

  • Robustness Evaluation: At 0 dB, CRB-guided perturbation reduces MPJPE from about 17.6 cm to 4.6 cm, an improvement of approximately 73.9%.The method also reduces MPJPE from about 5.8 cm to 4.0 cm at 30 dB.
  • Power Budget: The proposed scheme reduces low-power MPJPE to about 5.4 cm versus around 6.8 cm for fixed C = 1, an error reduction of approximately 21%.MPJPE generally decreases as the average power budget increases.
  • Sensing Duration: When sensing duration increases from 20 ms to 60 ms, the proposed scheme keeps MPJPE around 4.1–4.3 cm, about 33% below fixed C = 1.MPJPE increases with sensing duration because inference time decreases while inference frequency increases.
  • Communication Rate: Below 2.5 Gbps, the proposed scheme maintains MPJPE near 4.3 cm, reducing error by approximately 35% versus fixed C = 1.MPJPE increases with the minimum communication rate, especially in the high-rate region.
  • Inference Frequency: At Pmax = 1.5 W, the proposed scheme reduces MPJPE by about 14% versus fixed fin = 0.30 GHz and 32% versus fixed fin = 0.40 GHz.The two higher-frequency baselines become infeasible at Pmax = 1 W because of excessive computation energy.
  • Inference Depth: At Pmax = 2 W, the proposed method reduces MPJPE by about 33%, 15%, and 4% versus fixed C = 1, C = 2, and C = 3, respectively.Larger inference depth generally lowers MPJPE by extracting more temporal and spatial features.
Loading 2609.01908v1…