A Shared-Backbone Face Detection and Tracking Method for Edge Intelligence

Abstract

In edge computing scenarios, face detection and multi-object tracking (MOT) require efficient, low-latency real-time inference on resource-constrained devices. However, traditional methods typically employ independent detection and re-identification (Re-ID) models, leading to severe computational redundancy and poor real-time performance. To address these issues, this paper proposes a shared backbone face detection and tracking method for edge intelligence, which achieves collaborative optimization of detection and tracking tasks by reusing the backbone features of the detection network. Specifically, the detection model is constructed based on the lightweight network MobileNetV3, and efficient channel attention (ECA) and coordinate attention (CA) mechanisms are introduced to enhance feature representation capability. Meanwhile, an adaptive multi-scale feature fusion strategy is designed to generate discriminative Re-ID features, combined with a dynamic feature gallery update mechanism to improve tracking robustness. Finally, deployment verification is conducted on an embedded neural processing unit (NPU) platform, and mixed-precision quantization is adopted to optimize inference efficiency.Experimental results demonstrate that the proposed method achieves an excellent balance among accuracy, speed, and resource efficiency: the detection model parameters are reduced from 22.7M to 2.1M, achieving detection accuracies of 90.1% and 79.8% (Easy/Hard) on the WIDER FACE dataset, a MOTA of 75.2% and an IDF1 of 72.7% on the MOT17 dataset, and a real-time performance of 18.35 FPS on the Amlogic C302X chip platform. Compared with traditional decoupled schemes, the proposed method significantly reduces computational overhead while maintaining accuracy. This research provides an efficient and deployable face analysis solution for edge intelligence vision systems, which has significant application value in the fields of IoT and edge computing.