Research On Object Detection Algorithm In Real Traffic Scene Based On Deep Learning

The object detection in real traffic scene is the key to the construction of an intelligent transportation system,and it is also an important research topic.However,the complex real traffic environment,such as blocked vehicles,pixel overlap,weak light environment,and traffic lights and other small objects,will lead to the inaccuracy of object detection.These inaccurate detection results will directly affect the input information obtained of the intelligent transportation system,and then affect the control of the system on the traffic,which could lead to serious urban traffic paralysis.Therefore,how to quickly and accurately achieve the object detection in real traffic scene is the key and difficult issue we need to solve.To this end,this paper proposes different convolutional neural networks to solve the above problems.Aiming to detect the blocked vehicles efficiently in actual traffic scenes,we propose a multi-target corner pooling based neural network for vehicle detection.The hourglass network,which could extract local and global information of the vehicles in the images simultaneously,is chosen as the backbone network to provide vehicles' features.Instead of using the max-pooling layer,the proposed multi-target corner pooling(MTCP)layer is used to generate the vehicles'corners.And in order to complete the blocked corners that cannot be generated by MTCP,a novel matching corner method is adopted in the network.Therefore,the proposed network can detect blocked vehicles accurately.Experiments have demonstrated that the proposed network achieves an AP of 43.5%on MS COCO dataset and a precision of 93.6%on traffic videos,which outperforms the several existing detectors.In order to detect the vehicles in low light and pixel overlap,we propose a region attention-based neural network for vehicle segmentation.The hourglass network is used as the backbone network to generate feature maps of the input images,and these feature maps are sent to foreground(FG)and background(BG)branches,respectively.The FG branch is responsible for tasks related to things such as generating class,location,mask and the region of interest(Rol).The BG branch is responsible for the task related to stuff,which is semantic segmentation.The RoI from the FG branch is added to the BG branch through the proposed RoI attention module to provide object-level attentions.Finally,the feature maps of these two branches are sent to the stuff-things ranking module to generate the panoptic segmentation results.Experiments have demonstrated that the proposed network has strong robustness in complex real traffic scenes.And the proposed network achieves a PQ of 59.1%on the CityScapes dataset.Aiming at the problem that the convolutional neural network has a weak ability to detect small objects such as traffic lights and distant vehicles in real traffic scenes,this paper uses a feature guide attention module to enhance the network's ability to detect small objects.By extracting and processing the feature information of different convolutional layers,the effective features sent to the prediction module are increased,thereby enhancing the attention of the neural network to small objects.Experiment result shows that the APs on the MS COCO dataset is 28.6%,and the mAP on the Bosch dataset is 0.76,which performs better than other methods.