Traditionally, industrial robots have been programmed by highly specialized workers who either write code directly in platform-specific languages or use dedicated hardware (such as teach pendants) to guide the robot through specific via-points. However, the inherently complex and time-consuming nature of this process has hindered widespread adoption of industrial manipulators by companies that need flexibility to handle limited production volumes and rapidly changing product requirements. In response to this challenge, the introduction of walk-through programming approaches has significantly reduced both programming complexity and time, representing a crucial step toward meeting industry needs. Nevertheless, these solutions heavily rely on physical human-robot interaction, thus introducing the need to strike a balance between performance (i.e. robot speed during teaching) and safety of the human operator. To this regard, this paper proposes an architecture that integrates human motion perception with safety-oriented control algorithms in order to enable safe and efficient walk-through programming in actual industrial environments.