Is it possible to detect liquid with computer vision?

Yes, computer vision can detect liquids, but it depends on the context, the type of liquid, and the environmental conditions. The primary challenge lies in the physical properties of liquids—such as transparency, reflectivity, and dynamic shapes—which make them harder to identify compared to solid objects. However, by combining traditional image processing techniques with modern machine learning approaches, developers can build systems that detect liquids effectively in specific scenarios.

The most common methods involve analyzing visual features like color, texture, motion, or thermal signatures. For example, liquids like oil or colored fluids can be identified using color thresholding in HSV (Hue-Saturation-Value) space. Infrared or thermal cameras can detect liquids based on temperature differences, such as water on a road (which cools the surface) or hot coffee in a cup. Motion detection is useful for tracking flowing liquids, like a spill spreading on a floor, by analyzing changes between video frames. For transparent liquids (e.g., water), techniques like refraction analysis—observing distortions in background patterns—or using polarized light to reduce glare can improve detection. Machine learning models, such as CNNs (Convolutional Neural Networks), can also be trained on labeled datasets of liquid surfaces or spills to recognize subtle visual patterns like ripples or reflections.

However, real-world implementation requires addressing significant challenges. Transparent liquids often lack clear edges or contrast, making segmentation difficult. Reflective surfaces (e.g., glass, metal) can produce false positives by mimicking liquid-like highlights. Dynamic lighting conditions or occlusions (e.g., shadows, objects partially covering the liquid) further complicate detection. A practical solution is to combine computer vision with other sensors—like lidar for depth data or capacitive sensors for contact detection—to validate findings. For instance, in industrial settings, cameras might monitor liquid levels in tanks by detecting meniscus curves (the curved surface of the liquid), while ultrasonic sensors provide redundant measurements. In automotive applications, rain-sensing systems use infrared light and cameras to detect water droplets on windshields and trigger wipers. While no single approach works universally, domain-specific solutions leveraging multimodal data and iterative model training can achieve reliable results.