On-Road Object Detection Based On Deep Learning Methods

Intelligent transportation system is important for reducing traffic jams,ensuring traffic safety and improving traffic efficiency.On-road object detection is an important component of an intelligent transportation system.In addition,with the rapid development of deep learning,object detection based on deep neural networks has been widely accepted in recent years.In this thesis,some key problems in object detection are studied,such as detection of small objects,all-day detection and multi-class imbalance,details are as follows:(1)Based on Faster R-CNN(Faster Region-based Convolutional Neural Networks),a feature pyramid network is designed and implemented to detect traffic signs with small scales.In this network,coarse-grained and fine-grained features are used simultaneously.Experimental results show that the detection performance of traffic signs is greatly improved and better performance has been achieved compared with other previous methods.(2)Based on R-FCN(Region-based Fully Convolutional Networks),a new vehicle detection framework is developed for all-day detection of vehicles.In this framework,images are automatically categorized based on different lighting conditions.Experimental results show that the proposed framework can not only improve the vehicle detection performance,but also greatly save computational resources.(3)Multi-class imbalance is a tough problem when objects in different categories are detected in one neural network.In this thesis,data augmentation is applied and a hierarchical network architecture is developed.Experimental results show that the proposed methods can effectively improve the overall precision of various objects.To sum up,the developed methods in this thesis can effectively detect small objects in the images,improve the all-day detection of vehicles and reduce the effects of multi-class imbalance.The methods and the results can be used in future intelligent transportation systems for better on-road object detection.