Balanced Training for the End-to-End Autonomous Driving Model Based on Kernel Density Estimation

End-to-end autonomous driving models are widely used currently to avoid error transmission. The training of these models is affected by the imbalance of the training labels, such as imbalanced distributions of trajectory waypoint, steering, throttle, and brake labels. Therefore, the models cannot be trained enough on low-frequency labels. To mitigate this problem, a balanced training method is proposed in this paper. First, a general kernel density estimation method is used to estimate the probability function of labels. Then, to conduct balanced training, we design a cost-sensitive loss function that assigns different weights to each training data according to the probability of the labels. Comparative experiments on the Longest6 Benchmark show that the balanced training method proposed in this article improves the baseline significantly. We also conduct ablation studies to discuss the balanced training among different tasks in multi-task learning, as well as between two output forms: waypoint and action. Due to the generalization of kernel density estimation, the balanced training method proposed in this paper can be extended to any label distribution of two output forms in the end-to-end autonomous driving models.