Wood Vehicle Detection Based On Deep Learning

Felling trees without permission is a kind of illegal act that seriously threatens the forest resources.After the trees are cut down,they are transported by vehicles.It has certain concealment.Identifying the wood-carrying vehicles from the traffic monitoring in time can effectively stop the logging and theft,It is of great significance to protect forest resources.The target detection algorithm in Deep Learning is applied to the detection of wood vehicles,and the vehicle detection model is trained to identify the types of vehicles in traffic monitoring.Deep learning(DL)is a machine learning method based on neural network(NN)as the main modeling method.At present,it is widely used in image processing,text analysis,audio analysis and other fields.Among them,Convolutional Neural Networks(CNN)are mainly used in image processing,and target detection is an important field of research.Common target detection algorithms can be divided into one-step detection method and two-step detection method according to the existence of region proposal network,of which one-step detection method is represented by YOLO(You Only Look Once)series of algorithms,and two-step detection method is faster RCNN(faster Region Convolutional Neural Network)algorithm.In order to obtain a model with high detection accuracy,four algorithms such as Faster RCNN,Retina Net,YOLOv3,and Cascade RCNN(cascade Region Convolutional Neural Network)are first used to extract a network based on multiple CNN models to train some wooden vehicle detection models and compare theirs detection accuracy.And rate to analyze the applicable scenarios of each model.Considering the problems that may exist in practical applications,a Retina Net model with higher detection accuracy and faster speed is selected as the basic model for offline detection tasks,and a YOLOv3 model with high detection speed is selected as the basic model for real-time detection tasks.In order to speed up the model’s convergence and save the time cost required for training,a learning rate update strategy based on training loss is proposed for model training.The convergence trend of the model is judged by periodically determining the change trend of the loss,and updated accordingly Learning rate.For Retina Net,a GIo U(Generalized Intersection over Union)loss function is used as the bounding box regression loss function,and the form of focal loss(Focal Loss)is improved.For YOLOv3,the focal loss and GIo U loss function were used to improve its original cross entropy loss function(Cross Entroy,CE)and bounding box regression loss function to train a more accurate detection model.The experimental results show that the proposed improvement measures can further improve the average accuracy of the wood vehicle detection model.On the offline detection task,the average precision of the trained optimal Retina Net detection model for wooden truck detection is 97.7%,the mean average precision of the model is 80.4%,and the average inference time for detecting one picture is 75 milliseconds.The advantage of high precision can be applied to practical tasks as a reliable wooden vehicle detection model.On the real-time detection task,the trained YOLOv3 detection model has an average precision of 72.3% for the detection of wooden trucks,the mean average precision of the model is 62.8%,the detection rate is about 30 fps,and it can implement Real-time detection of wood-laden vehicles on surveillance video data with a frame rate below30 FPS,and certain detection accuracy can be obtained.