Research On Multi-target Detection Method In Traffic Scene Based On SSD

This paper studies the target detection methods for vehicles and pedestrians in traffic scenarios.The SSD target detection model is selected as the research object of this paper.The original intention of SSD model is designed for general data sets such as PASCAL VOC,MS COCO and ILSVRC.The difficulty in performing target detection on these data sets is that there are many types of targets to be detected,and the differences between different types of targets are large,and some types themselves have great deformation.The design of the SSD model structure is very good for detecting these targets.In the traffic scene,although the number of targets to be detected is small,and the deformation of the target itself is small,due to the particularity of the traffic scene,the targets to be detected are concentrated,and there is a common phenomenon of mutual occlusion,and the traffic field has a wide field of view.There are many smaller targets to be tested.Directly using SSD to perform target detection tasks in traffic scenarios,the effect is not satisfactory.In view of the difficulties in pedestrian and vehicle detection in traffic scenes,this paper improves the SSD model.The main work and innovations of this paper are as follows:(1)A convolutional neural network compression structure based on feature reuse is proposed.This method can reduce the parameter quantity and calculation of the model under the premise of ensuring the accuracy of the model,and then accelerate the target detection model.(2)A scale-invariant SSD(SISSD)algorithm is proposed.The theory of SSD detection algorithm is studied.The network default detection frame and the data enhancement during training are modified.The corresponding detection method is proposed,which makes the improved SISD model achieve better detection accuracy in vehicle and pedestrian detection.For the pedestrian and vehicle detection problems,this paper redesigned the SSD model structure and detection frame.The feature extraction network is improved by using feature multiplexing method,which reduces the parameter quantity and calculation amount of the feature extraction network,and adjusts the size and size of the default detection frame to make the detection frame and the real data distribution more consistent.Experiments were performed on the Kitti dataset.The improved SISD final model was improved by 13.13 compared to the original SSD average accuracy mean(mAP).The average detection accuracy(AP)of Car,Pedestrian,and Cyclist increased by 14.46,11.05,and 13.87,respectively.The improved model detection speed also reached 36.2fps.