Research And Implementation Of Vehicle Identification Technology Based On Deep Learning

With the improvement of the level of industrial automation and the advancement of science and technology,the application scenarios of object recognition are becoming more and more complex,and the requirements for the accuracy of object recognition are also increasing.The demand for vehicle identification in complex scenes is increasing day by day,and the difficulty of recognition is greater than that of general scenarios.The essence of deep learning is to train sample data by constructing neural networks of multiple hidden layers,and to learn features of samples automatically instead of relying on experience to extract features by hand.The most significant feature of deep learning is that it can split a complex problem into simple problems,firstly solving these simple problems in turn,and then integrating them to solve complex problems.Therefore,deep learning has become a key support for vehicle recognition and has its value for application.In practical applications,there are often cases where the target is severely occluded and the same target overlaps.In order to accurately identify the vehicle and accurately position it from a complex background,the YOLO(You Only Look Once)network model for deep learning is adopted.Based on the deep analysis of the YOLO(You Only Look Once)network model,this paper aims at solving the problem that the recognition rate of the original YOLO algorithm can not always meet the complex applications.Visual analysis of annotation files of the public datasets helped us to adjust the missing annotations and the mislabeled vehicle position is adjusted.And data augment is used to improve the vehicle recognition rate before a PCAR training dataset is formed;for false alarms can not avoid by current deep learning recognition algorithm when it comes to complex scenarios,a residual error-based algorithm is proposed.The false alarm suppression method of the first rule removes false alarms.The method of tuning the parameters of the deep learning model is studied.By adjusting the learning rate,the over-tuning phenomenon is caused by the large or little learning rate,which results in a slow learning cycle.By setting the weight decay coefficient,the excessive weight value can be avoided.Through comparison and analysis of training results of different training sets and verification sets on different networks,the experimental results show that when using the COCO dataset,the mean average precision(mAP)of the improved algorithm is 79.2%,which is 0.6% higher than that of the YOLO algorithm.When it comes to the PCAR dataset,the improved algorithm has an mAP value of 69.7.%,an increase of 7.4% over the YOLO algorithm.The improved algorithm makes the false alarm rate of vehicle recognition decrease and the recognition rate rise,which has certain theoretical and practical significance.