Convolution Neural Network with Selective Multi-Stage Feature Fusion: Case Study on Vehicle Rear Detection

Abstract

Vision-based vehicle detection is the most basic and important technology in advanced driver assistance systems. In this paper, we propose a vehicle detection framework using selective multi-stage features in convolutional neural networks (CNNs) to improve vehicle detection performance. A 10-layer CNN model was designed and visualization techniques were used to selectively extract features from the activation feature map, called selective multi-stage features. The proposed features contain characteristic vehicle image information and are more robust than traditional features against noise. We trained the AdaBoost algorithm using these features to implement a vehicle detector. The experimental results verified that the proposed vehicle detection framework exhibited better performance than previous frameworks.