Research On On-Road Object Detection Based On Deep Convolutional Neural Networks

Object detection system is a key component in areas such as ADAS(Advanced Driver Assistance System)and autonomous vehicles.It aims at detecting kinds of objects appearing in front of vehicles so it can give the driver warning information before accidents occurred.Object detection itself is a challenging task because objects often suffer from different gestures,uneven illumination or truncation.It becomes even harder when considering object detection in vehicle driving scenarios,because objects here are always small and the traffic condition can be very bad.In recent years,object detection based on deep convolutional neural networks have achieved great progress due to their powerful capability in both representation and learning,such as the RCNN(Region-based CNN)algorithm.Based on the Faster RCNN algorithm,this paper propose an accurate object detection algorithm used in vehicle driving scenarios,which comes from three efforts :(1)This paper design a new network aiming at detecting kinds of objects through merging multi-layer features and building ROI spatial pyramid pooling.(2)This paper digs prior information about the specific vehicle driving scenarios to promote detection time efficiency.(3)This paper uses online hard example mining to optimize the network training and adopts bounding box voting to optimize the detection results.Through the above measures,the proposed method can detect various objects in vehicle driving scenarios.Experiments on the KITTI dataset show that the proposed method performs better than the original Faster RCNN algorithm.