IK-SMOTE based data enhancement method for scene classification

In autonomous driving scenarios, the perception performance of perception cameras is one of the key factors in determining the performance and safety of autonomous driving. However, there are a variety of distractions during driving, such as plastic bags, mud and leaves. If the perception camera is occluded, the perception field of view of the perception camera will be greatly reduced, which can lead to traffic accidents. Therefore, the occlusion detection function of the perception camera is a key factor in ensuring the safety of autonomous driving. We propose a method called Inner K-means SMOTE (IK-SMOTE) to solve the problem of mismatching of occlusion scenes and normal driving scenes. This method significantly improves the performance of detecting model.