We evaluated the strategy performance on three different datasets (MNIST, FMNIST, and CIFAR10), which is simulated heterogeneity by assigning different data volume labels to these datasets. These datasets all consist of image data for vehicle perception tasks. The MNIST dataset contains 70,000 images from 10 different classes, including 60,000 train and 10,000 test samples . FMNIST and MNIST have similar data structures, both are grayscale images. In contrast, FMNIST focuses on more complex target recognition tasks, which contains 10 categories of everyday items .