Collision perception is a critical capability for intelligent mobile robots. Recent advancements in neural modeling have demonstrated that a composite neuron model, inspired by the Lobula Giant Movement Detectors (LGMDs) in locusts, offers enhanced collision selectivity by specifically responding to approaching targets. This characteristic shows great promise for improving robot collision detection as a sensor strategy. However, the model’s significant computational demands in terms of time and memory have hindered its application in robots with extremely limited computational resources. To address this problem, this paper introduces an optimized online collision perception visual system designed for a ground-based miniaturized robot. The system enhances the existing composite neuronal model by simplifying network structures and optimizing spatiotemporal functions. It has been implemented on a micro-robot equipped with an STM32F427 chip, with a diameter of approximately 4 cm and a weight of 50 g. The visual system efficiently processes continuous image streams at over 30 Hz while utilizing only 176 KB of memory. Arena tests have confirmed the system’s effectiveness and robustness in rapid collision detection, achieving a success ratio of 98.5%. Additionally, the results demonstrate the proposed model’s improved selectivity in handling high-speed motion stimuli, responding exclusively to approaching movements rather than receding or translating ones. This study offers a robust and parsimonious solution for mobile robots with constrained visual computing resources, effectively addressing the challenge of collision detection.