Energy-efficient traffic sign recognition using directly trained spiking neural networks and population decoding

Recognizing traffic signs is a fundamental perception task for automated driving systems and requires high accuracy under strict latency and energy constraints. Convolutional neural networks (CNNs) achieve strong performance but can be computationally demanding for embedded platforms. Spiking convolutional neural networks (SCNNs) offer an event-driven alternative that can reduce computation through sparse activity, yet their accuracy often degrades under very low-latency settings with few time steps. To improve spike-based inference under strict runtime constraints, we integrate a neural population decoding layer at the output stage and evaluate directly trained SCNNs with and without population decoding against a CNN baseline on the German Traffic Sign Recognition Benchmark (GTSRB). The best SCNN without population decoding achieved 98.85% test accuracy at 30 time steps, exceeding the CNN baseline of 98.38%. Population decoding improved performance in the low-latency regime, reaching 98.31% accuracy at a single time step, corresponding to an improvement of 0.56% over the SCNN without population decoding at the same temporal setting. Using an operation-based energy estimation, the SCNNs achieved over 14 times higher energy efficiency than the CNN at one time step. Overall, the results demonstrate that directly trained SCNNs can surpass a comparable CNN while enabling flexible trade-offs between accuracy, inference time, and energy efficiency. In particular, population decoding proves beneficial when operating under strict latency constraints.