The stride of a ConvNet tells you how much the kernels step when they slide (convolve) over the input image.
For example, with stride 1, at each convolution step they move one column to the right – and when they reach the end, they move down one row before starting moving right again:
With stride 2, at each convolution step your kernels move two columns to the right. When they reach the end, they also move two down.
Adding strides of > 1 to your ConvNet makes the pyramid generated by the convolutional downsampling more steep. It gives you a more efficient network (fewer parameters needed because the shape of your image gets smaller faster in your downstream layers) possibly at the cost of model performance.
It’s the trade-off between kernel size, number of kernels, and stride.