MobiAct: Efficient MAV Action Recognition Using MobileNetV4 with Contrastive Learning and Knowledge Distillation

TL;DR

MobiAct addresses the challenge of real time MAV action recognition under strict resource limits by combining a compact MobileNetV4 backbone with a teacher student knowledge transfer from ResNet18, a parameter free similarity aware attention mechanism, and a hybrid loss that blends contrastive and supervised objectives. The core innovations are Stage wise Orthogonal Knowledge Distillation across backbone stages, and a light attention module that enhances discrimination without adding parameters. Experimental results on three self collected MAV datasets show strong accuracy while maintaining low energy cost and fast inference, outperforming several state of the art while remaining suitable for on board deployment. The work demonstrates a practical path to reliable MAV visual communication and cooperative perception in swarm scenarios with limited computational budgets.

Abstract

Accurate and efficient recognition of Micro Air Vehicle (MAV) motion is essential for enabling real-time perception and coordination in autonomous aerial swarm. However, most existing approaches rely on large, computationally intensive models that are unsuitable for resource-limited MAV platforms, which results in a trade-off between recognition accuracy and inference speed. To address these challenges, this paper proposes a lightweight MAV action recognition framework, MobiAct, designed to achieve high accuracy with low computational cost. Specifically, MobiAct adopts MobileNetV4 as the backbone network and introduces a Stage-wise Orthogonal Knowledge Distillation (SOKD) strategy to effectively transfer MAV motion features from a teacher network (ResNet18) to a student network, thereby enhancing knowledge transfer efficiency. Furthermore, a parameter-free attention mechanism is integrated into the architecture to improve recognition accuracy without increasing model complexity. In addition, a hybrid loss training strategy is developed to combine multiple loss objectives, which ensures stable and robust optimization during training. Experimental results demonstrate that the proposed MobiAct achieves low-energy and low-computation MAV action recognition, while maintaining the fastest action decoding speed among compared methods. Across all three self-collected datasets, MobiAct achieves an average recognition accuracy of 92.12%, while consuming only 136.16 pJ of energy and processing recognition at a rate of 8.84 actions per second. Notably, MobiAct decodes actions up to 2 times faster than the leading method, with highly comparable recognition accuracy, highlighting its superior efficiency in MAV action recognition.

Figures (12)

• Figure 1: Schematic overview of our proposed model for visual communication: Actions serve as visual signals for inter-MAVs coordination, processed via a recognition model employing (1) stage-wise orthogonal knowledge distillation, (2) similarity-aware attention module, and (3) hybrid training strategy, enabling efficient and deployable MAV action understanding.
• Figure 2: Schematic diagram of the MobiAct architecture. Single-view or multi-view MAV action recognition methods, when executed on microprocessors, face challenges, such as latency and memory bottlenecks. To address the efficiency challenges in MAV action recognition, we leverage lightweight neural networks for high-performance deployment.
• Figure 3: MAV action category definition, including actions "start", "end", "1", "0".
• Figure 4: Overview of the proposed MobiAct framework. The MAV action video is processed in parallel by the teacher network (ResNet18) and the student network (MobileNetV4). Their high-level features are aligned through orthogonal projection and optimized using the Orthogonal Feature Alignment module. In parallel, a Similarity-Aware Attention (SAA) mechanism enhances discriminative feature representation. The student network is trained under a hybrid learning strategy guided by the teacher model.
• Figure 5: The overview of basic knowledge distillation. The student network generates projected features $Z$, while the teacher network produces normalized features $D$. The distillation loss is computed between $Z$ and $D$ to transfer knowledge from the teacher to the student.