Research On Lightweight Target Detection Algorithm In Road Scenes Based On Deep Learning

Road target detection is crucial in many research fields,such as advanced driving assistant systems and intelligent transportation.With the emergence of the high-performance computing platform and the development of deep learning technology,the detection accuracy of target detection algorithm has been dramatically improved.However,the scale of the trained model is large,and the use of the algorithm on mobile hardware devices is limited.The emergence of lightweight target detection algorithm makes it possible for target detection algorithms to be applied to intelligent terminal equipment,especially in onboard chips in autonomous driving.This thesis studies the problem that the lightweight target detection network has a simple structure and fewer parameters,leading to the algorithm’s accuracy being low.On the premise of satisfying the real-time performance and network lightweight,the detection accuracy of the algorithm is improved by adjusting the network structure of the algorithm.The main contents of the dissertation are:(1)Build a lightweight Mobilenet V3-SSD target detection network.The lightweight feature extraction network Mobilenet V3 is selected to replace the VGG16 network in the original SSD algorithm,and with the depthwise separable convolution in SSD algorithm instead of the conventional convolution further lightweight algorithm.At the same time,the construction of lightweight target detection algorithm is analyzed.(2)Aiming at the low efficiency of SSD algorithm feature map,a lightweight single-pole feature fusion mechanism was proposed.Three layers in the six-layer feature map are selected to perform a reverse fusion method that expands from the deep to the shallow layer,adding more semantic information to the high dimensional feature map.The improved network is verified on the filtered VOC data set,and the model effectively improves the detection accuracy of the lightweight target detection algorithm for targets in road scenes on the premise of only adding a few calculation parameters.(3)To solve the problem of low accuracy caused by large classification loss in lightweight target detection algorithm,a lightweight target detection algorithm using an asymmetric double detection heads for target position and classification was proposed.In the algorithm,the convolution head is used to detect the target position,and the group full connection head is used to detect the target classification.Only convolution layers is used to process the feature map in the position detection part.In classification detection,the feature map first passes through the convolution layer,fuses the feature map of the position regression branch,and finally uses the group full connection layer proposed in this paper to extract the classification information of the feature map.Category expansion was carried out on the VOC data set,and the electric tricycle category was added.The algorithm is then trained on the expanded VOC data set,and the results show that the classification loss of the improved model is significantly reduced,which effectively improves the detection accuracy of the lightweight target detection algorithm.Finally,the validity of the algorithm is verified on the KITTI data set.