Identity-Aware Attribute Recognition via Real-Time Distributed Inference in Mobile Edge Clouds

Zichuan Xu, Jiangkai Wu, Qiufen Xia, Pan Zhou, Jiankang Ren, Huizhi Liang

Abstract

With the development of deep learning technologies, attribute recognition and person re-identification (re-ID) have attracted extensive attention and achieved continuous improvement via executing computing-intensive deep neural networks in cloud datacenters. However, the datacenter deployment cannot meet the real-time requirement of attribute recognition and person re-ID, due to the prohibitive delay of backhaul networks and large data transmissions from cameras to datacenters. A feasible solution thus is to employ mobile edge clouds (MEC) within the proximity of cameras and enable distributed inference. In this paper, we design novel models for pedestrian attribute recognition with re-ID in an MEC-enabled camera monitoring system. We also investigate the problem of distributed inference in the MEC-enabled camera network. To this end, we first propose a novel inference framework with a set of distributed modules, by jointly considering the attribute recognition and person re-ID. We then devise a learning-based algorithm for the distributions of the modules of the proposed distributed inference framework, considering the dynamic MEC-enabled camera network with uncertainties. We finally evaluate the performance of the proposed algorithm by both simulations with real datasets and system implementation in a real testbed. Evaluation results show that the performance of the proposed algorithm with distributed inference framework is promising, by reaching the accuracies of attribute recognition and person identification up to 92.9% and 96.6% respectively, and significantly reducing the inference delay by at least 40.6% compared with existing methods.

Introduction

In modern smart city surveillance, two critical tasks are often performed in tandem: identifying individuals across different camera views (Person Re-identification) and describing their characteristics (Attribute Recognition). Traditionally, these tasks rely on cloud-based processing, which suffers from high latency, or on-device processing, which is limited by the low computational power of cameras.

Mobile Edge Computing (MEC) offers a middle ground by placing "cloudlets" near the data source. However, implementing real-time pedestrian tracking at the edge requires overcoming challenges related to model complexity and the inherent volatility of wireless networks. This paper presents a joint inference framework that optimizes both accuracy and latency through distributed processing.

Distributed Inference Framework

To maximize efficiency, the researchers developed a joint inference model that replaces two separate deep learning pipelines with a single, unified architecture. Based on a ResNet-50 backbone, the model extracts features useful for both identity and attribute recognition simultaneously. This multi-task approach reduces the overall model size by 55.2%.

The proposed distributed inference framework consists of four functional modules (A-D) that can be deployed across different edge nodes.

The system is partitioned into four distinct modules to facilitate distributed execution:

  • Module A (Detection): Executed on the camera to isolate pedestrian images, minimizing the data volume transmitted over the network.
  • Module B (Feature Extraction): Performed on an edge server to generate high-dimensional embeddings.
  • Module C (Distributed Gallery): Segments the identity database across multiple servers, allowing for parallelized similarity searches.
  • Module D (Fusion): Aggregates results to produce the final identification and attribute list.

Dynamic Module Distribution

Static assignment of these modules to servers is inefficient in dynamic environments where network traffic and server loads fluctuate. To address this, the paper introduces the ModDistMAB algorithm, based on a Contextual Multi-Armed Bandit (MAB) framework.

The ModDistMAB algorithm uses a contextual agent to dynamically choose the optimal server and link for module distribution based on real-time network feedback.

The ModDistMAB agent observes the current network context—such as delay levels and server availability—and learns the optimal placement strategy on the fly. By continuously updating its policy based on feedback (cost), the system minimizes end-to-end inference delay without requiring prior knowledge of network statistics.

Experimental Evaluation

The framework was evaluated using the Market-1501 and DukeMTMC-reID datasets, and implemented on a physical testbed comprising Raspberry Pis and Intel Neural Compute Sticks.

Key findings include:

  • High Accuracy: The system achieved 92.9% accuracy for attribute recognition and 96.6% for person identification. Training these tasks together proved synergistic, as identity and attributes provide mutual reinforcement.
  • Latency Reduction: Compared to fixed-assignment methods, the distributed approach and ModDistMAB algorithm reduced inference delay by over 40.6%.
  • Scalability: Parallelizing the gallery search (Module C) effectively removed the common bottleneck found in large-scale re-identification databases.

Conclusion

This research demonstrates that real-time pedestrian tracking is achievable at the edge by combining joint multi-task learning with adaptive distributed inference. By moving intelligence closer to the camera and dynamically managing workloads, the proposed system provides a scalable and responsive solution for next-generation surveillance in smart cities.

Summary updated

Abstract

With the development of deep learning technologies, attribute recognition and person re-identification (re-ID) have attracted extensive attention and achieved continuous improvement via executing computing-intensive deep neural networks in cloud datacenters. However, the datacenter deployment cannot meet the real-time requirement of attribute recognition and person re-ID, due to the prohibitive delay of backhaul networks and large data transmissions from cameras to datacenters. A feasible solution thus is to employ mobile edge clouds (MEC) within the proximity of cameras and enable distributed inference. In this paper, we design novel models for pedestrian attribute recognition with re-ID in an MEC-enabled camera monitoring system. We also investigate the problem of distributed inference in the MEC-enabled camera network. To this end, we first propose a novel inference framework with a set of distributed modules, by jointly considering the attribute recognition and person re-ID. We then devise a learning-based algorithm for the distributions of the modules of the proposed distributed inference framework, considering the dynamic MEC-enabled camera network with uncertainties. We finally evaluate the performance of the proposed algorithm by both simulations with real datasets and system implementation in a real testbed. Evaluation results show that the performance of the proposed algorithm with distributed inference framework is promising, by reaching the accuracies of attribute recognition and person identification up to 92.9% and 96.6% respectively, and significantly reducing the inference delay by at least 40.6% compared with existing methods.
BibTeX
@inproceedings{xu2020identity,
  author = {Zichuan Xu and Jiangkai Wu and Qiufen Xia and Pan Zhou and Jiankang Ren and Huizhi Liang},
  title = {{Identity-Aware Attribute Recognition via Real-Time Distributed Inference in Mobile Edge Clouds}},
  booktitle = {ACM Multimedia},
  year = {2020},
  url = {https://openreview.net/forum?id=BoZJKjU9dH},
}
Figure