Research On Small Traffic Light Detection Based On Lightweight Neural Network

Object detection is one of the hot research fields in deep learning today.In object detection,small target detection is extremely challenging because of its small target size and high detection difficulty.In order to detect small objects more accurately,experiments often require larger resolution images,but this also means greater computational overhead and slower image processing speed.In actual scene applications,the detection speed of the network It is also an important condition to consider,and traffic light detection is one of such application scenarios.At the same time,traffic light detection is often affected by environmental factors,such as occlusion,lighting,etc.,which adds another challenge to traffic light detection.In this paper,the traffic lights are divided into different groups according to the width and size.Based on this,a layered optimization method for a priori anchor design is proposed.The number and size of independent a priori frames are designed for each group without affecting each other.It solves the problem that the shape of the prior anchors and the traffic lights are insufficiently matched so that prior anchors cannot effectively characterize small traffic lights,and effectively improve the detection effect of small targets in traffic lights.In this paper,an adaptive network architecture design and training method are designed for the situation of extremely small traffic lights.By selecting a larger resolution feature map to improve the detection rate of extremely small traffic lights,at the same time,by adding a feature enhancement module to solve the problem The problem of insufficient feature expression ability and too shallow depth of the size feature map effectively improves the detection effect of extremely small traffic lights.In this paper,combined with the characteristic that the shape of the traffic lights are mostly vertical rectangles,the corresponding convolution kernel structure and pooling structure are designed,which additionally reduces the size of the feature map during prediction,while effectively reducing the network parameters and the amount of calculation.,Slightly improved the detection effect of the network.The data set selected in the experiment of this paper is the Bosch small traffic light data set.The traffic lights with a width of less than 4 pixels after the experiment scaled down accounted for 20%of the total,which is more challenging than the LISA and other data sets commonly used in previous work.The data set still has occlusion,environmental interference and other factors.The existing traffic light detection results for this data set are not very satisfactory.In this paper,the mAP of this data set has reached 69.4%,and the single sheet detection speed is 7.016 milliseconds.It is superior to YOLOv3,SSD and other commonly used lightweight detection networks in accuracy and speed,and has good application prospects in applications such as autonomous vehicles.