Spatio-Temporal Pruning: A Generic Method for Edge-Based 3D Distracted Driver Recognition

Abstract

Advanced 3D models such as 3D convolutional neural networks (CNNs) and vision transformers (ViTs) can accurately recognize distracted driver behavior in videos. However, these models demand considerable computational resources, making real-time inference on embedded platforms challenging. To address this issue, we propose a Generic Spatio-Temporal Pruning (GSTP) framework for edge-based distracted driver recognition that is plug-and-play and compatible with heterogeneous video backbones, including both 3D CNNs and ViT-based models. The core idea of GSTP is to reduce spatial and temporal redundancy by leveraging the characteristics of in-cabin distracted driver recognition. For spatial pruning, we present a compare-and-select strategy that first extracts motion cues from adjacent frames and then selects informative spatial regions through shifted windows on the motion map. For temporal pruning, a score-and-prune module is introduced to evaluate temporal redundancy and remove less informative temporal features. In this way, GSTP improves the efficiency of video backbones for edge deployment while maintaining or improving recognition performance. Experimental results on the 100-Driver and 3MDAD datasets demonstrate that backbones equipped with GSTP achieve faster inference on edge platforms (Nvidia Jetson TX2 and AGX) and competitive or improved accuracy with reduced computational cost across different backbone architectures