An acquisitive knowledge distillation method for deployment at the edge of the road

With the rapid development of intelligent transportation systems, road-side units (RSUs) face significant challenges in traffic perception tasks: the computational demands of deployment models often exceed the resources available on edge devices. Knowledge distillation (KD) is commonly employed to facilitate edge deployment. However, the issue of knowledge mismatch between teacher and student models in KD limits the performance of deployed models. We propose a Knowledge Acquisition-oriented Knowledge Distillation (KAKD) method for road-side edge devices. This method employs a bidirectional knowledge interaction mechanism, enabling the student model to actively filter and assign importance to the required knowledge, effectively addressing the core issues of redundant knowledge interference and insufficient acquisition of critical knowledge in traditional KD. KAKD introduces and improves the PID control algorithm to construct a dynamic feedback mechanism, quantifying knowledge importance from three dimensions: loss value, sample difficulty, and knowledge forgetting value. It enhances the analysis of historical loss trends through n-th order difference and amplifies the differences between samples using the σ function. Additionally, we design a Sparse-softmax activation function with sparse probability distribution to focus on high-value knowledge. Finally, we construct a traffic scene dataset covering 7 cities and 5000 images to effectively validate the robustness of the method. Experiments show that KAKD outperforms peer methods across multiple datasets. When combined with YOLOv8n and compared to the base YOLOv8n model, the mAP50 value increases from 77.84% to 79.01%, a 1.17% improvement, and the mAP50-95 increases from 49.43% to 50.2%, a 0.77% improvement.