Research On Lightweight Binary Network In Traffic Scene

There are a large number of environmental perception requirements in traffic scene.Whether it is in driver assistance or automatic driving,it is usually necessary to detect and identify lane lines,drivable areas,vehicles,pedestrians,road signs,traffic lights,traffic signs and other important traffic targets,and the above-mentioned important information is detected and identified from the environment through visual sensors,which can help vehicles make safe and standardized driving behavior decisions.At present,with the success of deep neural networks in the field of computer vision,various detection and recognition tasks in traffic scenes can be processed by deep neural networks.The computing power and memory requirements are high,and the energy consumption is also high,making it difficult to deploy on resource-constrained platforms.As an intelligent vehicle,its on-board platform is constrained by limited resources,and also requires low power consumption and strong real-time performance.Therefore,it is of great significance and practical to study lightweight object detection or image classification neural networks for resource-constrained on-board terminals.Aiming at the problems of vehicle and pedestrian detection and traffic sign recognition in traffic scenes,based on the full-precision deep neural network model,we make a research on the binary neural network about object detection and image classification.The specific contents include:The binarized SSD network for vehicle and pedestrian detection is studied.First,the convolutional unit is binarized by the binary unit with shortcut branch for SSD300 with VGG16 as the backbone network,and the influence of the magnitude-aware factor of the weight is considered when binarizing the weights.Then set the appropriate scale and ratio of bounding boxes through the clustering algorithm,and finally replace the STE with the derivative of the hyperbolic tangent function to approximate the derivative of the Sign for backpropagation.Finally,the real-valued network and the binarized detection network of this paper are compared on the KITTI and PASCAL VOC datasets respectively.Under the premise that the accuracy is not greatly reduced,the computational cost is reduced to six percent of the original,and the amount of parameters is reduced to fifteen percent of the original for KITTI dataset;the computational cost is reduced to eight percent of the original,and the amount of parameters is reduced to twenty-two percent of the original for PASCAL VOC dataset.The binary ShuffleNetV2 network for traffic sign recognition is studied.Firstly,the convolution units except the first layer are binarized on the basis of ensuring that the convolution mode of the usual convolution and Depth Wise convolution of Shuffle Netv2 remains unchanged.At the same time,it is observed that in the binary network,the activation distribution is closely related to the semantic feature representation of the feature map,and the change of the semantic feature representation of the feature map will greatly affect the accuracy of the network.Therefore,by introducing learnable parameters into the Sign and activation function to adjust the activation distribution,the expression ability of the network is improved,and then the accuracy of the network is improved.Finally,the real valued network and the binary recognition network in this paper are compared on GTSRB and BTSC datasets.When training BTSC datasets,considering that there are too few samples of BTSC datasets,the features learned from GTSRB datasets are transferred to BTSC datasets.The recognition model in this paper has achieved good results in computational cost,model size and recognition accuracy.On the premise of no significant decrease in accuracy,the computational cost is reduced to eleven percent of original and the amount of parameters is reduced to nine percent of original for GTSRB dataset;the computational cost is reduced to eleven percent of original and the amount of parameters is reduced to eleven percent of original for BTSC dataset.