A
monocular image-based three-dimensional lane line detection method, device and medium, two-dimensional semantic features are extracted from a
monocular lane image, and three-dimensional
perception features are obtained by using a pre-trained deep model; two-dimensional semantic features and three-dimensional
perception features are fused, an initial lane query is enhanced, and a query representation for three-dimensional lane line prediction is obtained; the query representation is iteratively decoded layer by layer, the sampling offset in the decoding process is dynamically adjusted based on the uncertainty parameters obtained by decoding, and the two-dimensional semantic sampling values and three-dimensional
perception sampling values obtained by sampling in the decoding process are adaptively weighted and fused to obtain a three-dimensional lane line prediction result; the detection network is trained and constrained based on geometric structure constraints, and the final three-dimensional lane line detection result is output. The present application improves the
spatial perception ability, prediction stability and lane structure rationality of
monocular three-dimensional lane line detection in complex road scenes without relying on additional depth sensors and depth labeling information.