| In advanced driver assistance systems and intelligent transportation systems,the detection of road objects,including vehicles and pedestrians,is crucial.Detection of road targets through visual images that have different backgrouonds is very difficult.Due to the requirement of the automatic driving system and the intelligent traffic system,the road object detection algorithm needs to have a higher accuracy while at the same time have a higher operating speed on the embedded device.Therefore,it is of great practical value to develop road object detection algorithm that can achieve higher running speed on embedded devices.Convolutional neural network has achieved excellent results in the field of image object detection.Therefore,using convolutional neural network to achieve road object detection was of great research value.Object detection using convolutional neural networks has also become a research hotspot.This paper aims at the characteristics of road object detection,reconstructs the SSD network,improves the detection speed and detection accuracy of road targets,and conducts experiments on embedded simulation platforms and automotive embedded platforms.The main work of this paper is as follows:(1)Through the research of the SSD object detection network,the MobileNet network was used to build the MobileNet-SSD(MSSD)network,which reduced the number of parameters and calculations of the SSD network and accelerated the SSD object detection process.(2)Improved the MSSD network for road objects and changed the aspect ratio and scale factor of the anchor box in the MSSD network.The detection accuracy of the new MSSD network for human object was improved by 3.32(AP)compared to the SSD network with twice the speed.(3)A sub-pixel anchor-box network(SPMSSD)is proposed.By arranging extra anchor box between the anchor point,the overlapping area between the MSSD anchor box is improved,and the MSSD anchor box has a better match to the real object.The hit rate of the box improves the accuracy of the MSSD object detection network.The SPMSSD has doubled the detection speed of the SSD while improving the detection accuracy on the road object by 8.42(mAP)and the small road target by 7.45(mAP).(4)The SPMSSD network was implemented on the embedded device NVIDIA Jetson TX2,offline simulation was completed on the self-built data set,and on-board road tests were conducted.The SPMSSD network can run stably at twice the speed of the SSD network on the on-vehicle platform and achieve higher detection accuracy. |
