SEdgeNet: Using millimetre-wave radar and stochastic graph operators to recognise activities without cameras

# Why radar instead of cameras

# The technical challenge

# What SEdgeNet changes The Aston University team (Vincent Gbouna Zakka, Luis J. Manso, Zhuangzhuang Dai) rethink a common graph-based operator, EdgeConv, used in point-cloud networks. EdgeConv builds a local graph by connecting each point to its k nearest neighbours and computes features along those edges. That deterministic use of all k neighbours is useful for clean data but brittle for radar: the same noisy neighbours repeatedly corrupt features during training.

SEdgeNet introduces two linked mechanisms under the name Stochastic Edge Convolution (SEdgeConv):

• Edge-wise Factorised Convolution (EFC): conventional EdgeConv applies dense transformations across edge feature channels, which is costly and ill-suited to dynamically varying graphs. EFC adapts depthwise separable ideas to irregular edge features: each channel is transformed independently, then per-channel responses are fused with a pointwise projection. The factorisation lowers computation when combined with stochastic neighbourhoods.

Layer outputs use symmetric max pooling to aggregate neighbour features into permutation-invariant point representations. SEdgeNet stacks multiple SEdgeConv layers and concatenates their outputs along channels to build a multi-scale embedding that preserves local geometry and more abstract structure.

# Why the design choices matter Stochastic sampling acts like a regulariser specific to graph neighbourhoods: the network cannot depend on fixed local patterns that might be artifacts of noise or multipath. The factorised edge processing reduces the computational burden that otherwise grows when neighbourhoods change between passes. Together these mechanisms tailor graph-based point-cloud learning to the sparse, irregular character of radar returns.

# Practical implications SEdgeNet aims to enable cameraless sensing for tasks where privacy or lighting are concerns: fall detection in homes or care settings, gesture interfaces for smart devices, and other human activity recognition tasks. Using radar point clouds avoids recording images while still delivering spatial and motion cues.

# What remains open The article describes the architecture and the motivation but does not present performance numbers or deployment details in the supplied text. Questions about robustness across different radar hardware, real-world environments, energy and latency on embedded devices, and integration with privacy policies remain to be evaluated in follow-up work.