The perceptron algorithm was invented in 1957 at the Cornell Aeronautical Laboratory by Frank Rosenblatt (Wikipedia EN)
It was the first neural network like architecture. It was thought that (a combination of) perceptrons could learn anything. There was a lot of controversy about this and after a while Marvin Minsky showed that a single perceptron could not solve the XOR problem. This was the end of the first neural network hype.
After a couple of decades the multilayer perceptron was invented, and with it the backpropagation algorithm. The backpropagation algorithm can learn weights that are not directly connected to the output unit. However, it needs a differentiable activation function. Hence, the ramp activation function was not sufficient anymore and others like the sigmoid and tanh were used for this.
Besides, the perceptron was first implemented using electronic boards and not with the software available to us today. Binary thresholds made this complicated process a lot easier.
So, in short Rosenblatt could have chosen for any version of the heaviside step function but he chose for the one where $w \cdot x + b = 0$ leads to a zero output. If you think binary values this makes sense, a zero input (i.e. False) leads to a zero output.