Research On Target Localization Based On Neural Network

With the development of technology,the demand for multi-scene,anti-interference and realtime target positioning algorithms in communication,radar,traffic and other fields is becoming more and more urgent,and the development of neural network technology in recent years has made it possible.The algorithms of direction finding based on fully connected neural networks and the algorithms of license plate detection based on convolutional neural networks are studied in this paper.The task of direction finding is to determine the location of the signal sources based on the array received signal.The task of license plate detection is to find the location of the license plate in the picture and determine its size.Traditional direction of arrivals(DOA)estimation algorithms and license plate detection algorithms have many problems such as poor robust of factors and limited application scenarios.When the scene faced by the traditional algorithms changes,the algorithms must be improved for the specific scenario before they can continue to be used.In addition,when the environmental factors in the scene change,the performance of the traditional algorithms will be seriously affected,and may even fail.Based on this,the neural network is introduced into the two positioning problems of DOA estimation and license plate detection in this paper,and the experimental results show that the performance of new algorithms is excellent.In this paper,the neural network is introduced into the direction finding problem,and a direction finding algorithm based on neural network is proposed.The experimental results show that the algorithm successfully solves the problem of direction finding in the case of sources coherence and array with position error.And the algorithm runs faster than traditional algorithms,and can solve the direction finding problem in real time.In this paper,the neural network is introduced into the license plate detection problem,and the license plate detection algorithms based on YOLOv3,SE-YOLOv3 and L-YOLOv3 are proposed.The experimental results show that the algorithms have excellent speed and accuracy and are robust to disturbance factors such as illumination changes,shooting angle changes,weather interference,and smear.Based on the in-depth study of the working principle of YOLOv3,the anchors and network structure of YOLOv3 are modified according to the characteristics of this task in this paper.Finally,YOLOv3 is successfully applied to the license plate detection task.The algorithm is then tested on a test dataset that contains six scenarios.The experimental results show that the accuracy of the algorithm is far more than the traditional algorithms.In addition,the experimental results show that the proposed algorithm is robust to environmental factors and can detect license plates in various interference situations.For the problem that the positioning results of the license plate detection algorithm based on YOLOv3 is not accurate enough and the error between the actual position of the license plate is slightly larger,the channel attention mechanism is introduced into YOLOv3 and then SEYOLOv3 is proposed in this paper.The experimental results show that the positioning result of SE-YOLOv3-based algorithm is more accurate than YOLOv3-based algorithm.L-YOLOv3 is proposed in this paper for the problems of YOLOv3-based license plate detection algorithm such as too many parameters,too large model weight file,and too much detection time.L-YOLOv3 uses a new,more lightweight feature extraction network and a new redesigned detection network based on the task requirements.In addition,L-YOLOv3 reduces the model input scale and the detection feature scale.The above improvements greatly reduce the parameters,calculations,the size of weight files and the detected time of L-YOLOv3.In addition,the positioning results of L-YOLOv3-based algorithm are more accurate than YOLOv3-based algorithm.