Mixed Algorithm Real Time Image Processing for Robotic Drones in Deterministic Environments

This paper presents a mixed algorithm approach for real-time image processing on air and land based robotic drones that operate in highly deterministic environments such as automated warehouses. A combination of heuristic, multi state algorithmic solution based on Canny Edge and a CNN deep learning-based solution is proposed to ensure maximum performance out of low power embedded hardware. A heuristic i.e., rule-based approach is used for fixed, simple tasks such as navigation, docking, charging, identifying racks etc. On detecting an edge case or an anomaly beyond the scope of heuristics, the program switches to a deep learning-based approach in real time for extracting context and navigating the complex scenario. This computationally intense mode only lasts for short bursts and switches back to an appropriate, power efficient algorithm. The paper aims to study the effects of different embedded computer vision methodologies taking into account the context switching time and their effect on the computational resources such as CPU, RAM and ROM along with variations in CPU temperature and verify the effectiveness of a mixed algorithm approach on a Raspberry Pi 4 single board computer running “Raspbian Bookworm” a 64bit Debian based operating system interfaced with a CSI Camera for visual feed.