Research On Convolution-Neural-Network-Based Detection Algorithms For Traffic Light

With the rapid development of economy,there is an increasing number of cars on the roads and streets.This phenomenon will result in a series of problems,such as the traffic jams and frequent traffic accidents.In order to solve these problems,intelligent transportation system has been widely used in road traffic,and plays a very important role.In an intelligence traffic system,the traffic light detection is an important point.However,due to the road scenario is complex and changeable,and the size of traffic lights in such situation are smaller than the other objects,the traffic light detection is still difficult to be solved in the actual scene.Recently,based on the powerful ability of feature learning,convolutional neural network has been widely exploited in the fields of image classification and object detection,and made a series of achievements.Inspired by the appealing superiorities,this paper carries out a comprehensive investigation on the traffic light detection（i.e.,the small object detection）with the help of the convolutional neural network.The main networks and contributions of this paper lie in:Firstly,based on the Faster R-CNN object detection algorithm,we propose an improved design method for the anchor.This design method can utilize smaller anchors compared with the conventional schemes.As a consequence,a proposal region,which can suitable for the size of traffic lights,can be superiorly generated in the region proposal network（RPN）.Furthermore,according to the characteristics of the small object detection（i.e.,the detailed information about the small object is most stored in the shallow feature maps of the Faster R-CNN network and its corresponding rich semantic information is most stored in the deep feature maps of the Faster R-CNN network）,we propose a feature fusion strategy.Specifically,the Conv5₃ feature map in the Faster R-CNN feature extraction network is first processed by a specific deconvoluting operation.Then,such feature map is further processed by the feature fusion with Conv4₃ feature map.In this fusion operation process,we propose two novel fusion methods,referred to as Eltwise（PROD）and Concat respectively.The Conv45 feature map generated by the feature fusion operation is considered as the last layer in the feature extraction network and transmitted into the RPN and the region of interest pooling（ROI Pooling）.Finally,combing the above mentioned two feature fusion methods and the improved design method for the anchor respectively,we propose two improved algorithms,referred to as E-Faster RCNN and C-Faster RCNN.Secondly,based on the SSD object detection algorithm,a Conv3₃ high resolution feature map is merged into the multi-scale convolution detection.It is noteworthy that the Conv3₃ and Conv4₃ feature maps can generate a lot of prior boxes that match the size of small object,but the Conv3₃ and Conv4₃ feature maps are the shallow feature maps,which lack some rich semantic information.Thus,we propose a novel method,referred to as multi-level feature fusion.Specifically,the Conv5₃ deep semantic feature map,which is first processed by a deconvoluting operation,is merged into the Conv4₃ feature map.Then,the fused feature map will be further processed by a deconvoluting opeartion and merged into the Conv3₃ feature map.Furthermore,based on each feature layer of the SSD network feature maps,the changing process of detailed information about the traffic light is also analyzed in this paper.Analysis results show that the Conv10₂ and Conv11₂ feature maps will not work with the traffic light detection.In order to further reduce the computational load of the network,two pairs of convolution layers（i.e.,the Conv10₁/Conv10₂ convolution layer and Conv11₁/Conv11₂ convolution layer）can be removed and thus an improved algorithm,called MF-SSD,is generated.Finally,based on the American LISA traffic light detection dataset,we make a comprehensive experiment for the E-Faster RCNN,C-Faster RCNN,and MF-SSD algorithms respectively.Experimental results show that the mAPs of E-Faster RCNN and C-Faster R-CNN are 84.64% and 83.99%,which are 11.72%,11.07%,and 3.86%,3.21% higher than that of Faster R-CNN and LOCO algorithms,respectively;The mAP of MF-SSD algorithm is 86.45%,which is 4.04% and 5.67% higher than that of SSD and LOCO algorithms,respectively.Thus,experimental results reveal that the improved E-Faster RCNN,C-Faster R-CNN and MF-SSD algorithms possess desirable detection effect and achieve accurate traffic light detection.