Control method for self-moving device and self-moving device

The patent replaces expensive, complex hardware (GPU+NPU) with a simpler, more affordable processor configuration that can still perform image identification tasks. The solution uses a lightweight neural network model that runs on standard microcontrollers or low-power processors, making the system more accessible and cost-effective while maintaining adequate identification accuracy for practical applications.

The patent transforms the image identification system by changing the parameters of the neural network model - reducing model complexity, decreasing the number of layers and parameters, and optimizing for embedded deployment. This parameter transformation allows the system to achieve acceptable identification accuracy with significantly reduced computational requirements, resolving the contradiction between accuracy and hardware demands.

The patent extracts and removes the requirement for specialized processing units (GPU and NPU) from the system architecture. By designing a lightweight neural network that can execute on standard processors, the solution eliminates the need for these complex, expensive components while preserving the core functionality of object identification, thus improving productivity without increasing device complexity.

The patent substitutes the mechanical/hardware-based solution (adding GPU and NPU components) with a software-based solution (optimized lightweight neural network). This substitution replaces the need for additional physical processing units with algorithmic optimizations and efficient code implementation, achieving better identification performance without increasing hardware complexity.

The patent applies partial action by implementing a neural network with just enough complexity to achieve practical identification accuracy, rather than using a full-scale high-performance model. This partial implementation provides sufficient computing speed for real-time applications while consuming significantly less power, as the model processes images with fewer computational operations and smaller data throughput.

The patent introduces dynamic adaptivity into the system by implementing adjustable inference parameters, optional preprocessing steps, and configurable model variants. This allows the computing speed and power consumption to be dynamically balanced based on operational requirements, enabling the system to operate efficiently across different scenarios without requiring always-maximal processing power.