Vulnerable Road Users Detection using Convolutional Deep Feedforward Network

Abstract

A new convolutional deep feedforward network (C-DFN) is proposed to detect vulnerable road users at 57.9% misclassification rate using Caltech Dataset. Instead of going deeper, three C-DFN is stacked to achieve 43.4% misclassification rate. Part-based C-DFN further reduces the rate of 42.5% to tackle occlusion problem. In addition, investigation of adaptive activation functions are performed to understand the effect of saturated and non-saturated functions in mitigating the vanishing and exploding gradient issues of neural networks.