Flexible neural network for fast and accurate road scene perception

Abstract

Accurate object detection on the road is the most important requirement of autonomous vehicles. Extensive work has been accomplished for car, pedestrian, and cyclist detection; however, comparatively, very few efforts have been put into 2D object detection. In this article, a dynamic approach is investigated to design a perfect unified neural network that could achieve the best results based on our available hardware. The proposed architecture is based on CSPNet for feature extraction in an end-to-end way. The net extracts visual features by using backbone subnet, visual object detection is based on a feature pyramid network (FPN). In order to increase the net flexibility, an auto-anchor generating method is applied to the detection layer that makes the net suitable for any datasets. For fine-tuning the net, activation, optimization, and loss functions are considered along with multiple check points. The proposed net is trained and tested based on the benchmark KITTI dataset. Our extensive experiments show that the proposed model for visual object detection is superior to others, where other nets output very low accuracy for pedestrian and cyclist detection, our proposed model achieves 99.3% recall rate based on our dataset.