ResCoNN: Resource-Efficient FPGA-Accelerated CNN for Traffic Sign Classification

Abstract

Precise classification and detection of (distorted and normal) traffic signs in real-time is one of the non-trivial requirements for safe autonomous driving. The state-of-the-art convolutional neural networks (CNNs) for traffic sign detection though accurate are resource-hungry due to their inherent structure with massive networks with millions of full-precision parameters making them infeasible for low-end FPGA platforms leading to higher implementation costs. The existing works employing low-precision bits and similar techniques though hardware friendly leads to degradation in accuracy. In this work, we propose a resource-efficient CNN (ResCoNN) architecture with a small number of weights (only 60,000 compared to a few million in state-of-the-art CNNs) and employ for traffic sign detection and classification. For increased efficiency, the network takes advantage of binary weights and integer activations, rather than employing complex computations like Batch Normalization and Exponential Linear Units. ResCoNN achieves a classification accuracy of > 96% on real-world images at a framerate of 36fps on a Zynq SoC (xc7z020clg484-1) with 90% reduced weights compared to state-of-the-art CNNs.