Study On Pseudo-label Based Domain Adaptation Of Object Detection

The car environment perception system can provide information support for the subsequent decision-making of autonomous driving,and the stability and accuracy of the perceived information are related to the safety of the car,and it is the safety guarantee for the realization of autonomous vehicles(AVs).Image object detection based on Convolutional Neural Network(CNN)is currently widely used in the environment perception task of AVs.Although this type of algorithm has the advantages of accuracy and efficiency compared with traditional methods,it also has the following obvious shortcomings: 1)As the data-driven algorithms,its require massive annotation data to support,which brings time-consuming,laborious,and costly data collection and data annotation burden;2)The algorithms assume that the training set and the test set have the same feature distribution,which is usually not satisfied in autonomous driving applications.The domain shift will appear when the feature distribution is different,and this will cause a significant decrease in the performance of the detector.Domain adaptive learning,as a branch of transfer learning,can use existing knowledge to solve problems in related fields,and can effectively improve the dependence of CNN on labeled data.Therefore,this paper uses a pseudo-label-based domain adaptive method to adapt the detection accuracy of the detector in a variety of weather and light conditions or in a variety of urban environments.The main works are as follows:(1)Pseudo label based domain adaptation framework.First,the pseudo-labels of the current driving environment(target domain)are generated by the front-end detector and optimized by the data optimization module to correct errors.The adaptive sampling module is approached to sample target domain image according to the number of hard samples per image to select more effective data.Finally,to reduce the influence of noise information,knowledge distillation and label smoothing are used in the retraining module to update the front-end detector more robustly.This paper uses several public datasets and a local driving video to do the cross-domain experiment and compares it with the current advanced algorithms.The method shows better performance than existing methods.(2)Pseudo label based online domain adaptation framework.Based on the previous work,this paper proposes an online domain adaptation framework for the online input requirements of the target domain.First,the front-end detector generates pseudo-labels of continuous online input images.Then the pseudo-label optimization module uses the source domain and target domain data for adversarial training to correct the pseudo-label errors.The framework uses the online input images to iterative update the front-end detector.This paper uses BDD100 K to do the cross-domain experiment,the results show the detector performance has significantly improved by iteratively updating.