Object Detection In Driving Video Based On Deep Learning

In recent years,autonomous driving technology has attracted widespread attention.Selfdriving vehicles need to perform real-time object detection through driving video,and then make timely decisions during driving.Therefore,a robust object detection algorithm for driving video has important practical significance.The generalization ability of traditional detection algorithms is not good enough,and most of them cannot adapt to the complex and changing environment and real-time requirements.The use of deep learning for object detection can avoid many problems of traditional algorithms and obtain detection results far superior to traditional methods.At present,most object detection during driving videos is only for vehicles.In fact,information such as traffic lights,pedestrians,and riders is also necessary for environmental perception.Therefore,this paper uses a deep learning-based object detection algorithm and improves its feature extractor,which is applied to object detection of multiple categories under driving video.The main work of this paper is as follows:Aiming at the shortcomings of traditional algorithms such as poor robustness,low accuracy and poor real-time performance,this paper uses deep learning algorithm to construct multi-object detection model of driving video,and improve detection accuracy.The YOLOv2 algorithm has good real-time performance and high detection accuracy.The deeper network layer number helps the feature extractor in the algorithm to extract features.Based on the YOLOv2 algorithm,this paper replaces the backbone of the darknet19 neural network with a 53-layer network with residual blocks,and obtains higher detection accuracy by increasing the number of network layers of the feature extractor to achieve better detection results.A series of verification experiments and applications were performed on the improved algorithm.Configure a deep learning framework such as TensorFlow and CUDA computing architecture on a PC with GPU;process KITTI object detection dataset and Udacity vehicle detection dataset;use CIFAR-10 dataset to evaluate the classification ability of the improved algorithm,and improve the neural network in this paper.The network in this paper and the darknet19 and VGG16 networks respectively train and compare the classification performance.The accuracy and recall rate of the improved model in this paper are 0.934 and 0.872 respectively,and both indicators are higher than the latter two networks;the KITTI dataset is used to evaluate the improved algorithm.a total of 1106 test samples participated in the evaluation,the average accuracy of the model obtained by the training was 75.36%,and the mAP obtained by the integrated pedestrian was 65.27%,which was 4 percentage points higher than the model trained by the original YOLOv2 algorithm;the training has more categories.And more small objects of the udacity data set for driving video object detection,the overall effect is better than the original YOLOv2 algorithm training model.Finally,using the training of this paper to get the model for the actual video test has got a good detection effect.The experimental results show that the improved feature extractor(neural network)is helpful to extract features.The improved YOLOv2 algorithm has higher detection accuracy than the original YOLOv2 algorithm,and it is very good for object detection under driving video.