Scene Recognition TRIZ Cases

Class-independent objectness loss trains a neural network to detect unknown objects while preserving known-class accuracy.

Manual 3D annotation of traffic management objects is slow and error-prone; multi-frame 2D boxes localize 3D boxes and generate training cutouts.

A vision-based model projects image features into adjustable virtual camera space to detect and classify road users without radar or Lidar.

Vehicle cameras capture peripheral images with position data, allowing schools to receive alerts when students leave school routes without mobile terminals.

A gap classifier checks interrupted rail sequences at turnouts to improve image-based rail path identification in changing weather.

Four detected slot edges and assumed rectangle geometry enable accurate pose estimation from 2D images with lower computational demand.

Characteristic scale distributions and receptive-field sizes guide shared CNN architecture selection, reducing design effort for embedded multi-task vision.

Recurrent encoder-decoder passes combine candidate-area context to classify activated vehicle lamps from a single image with lower computational demand.

Annotated datasets identify robot-reachable regions and relevance values to assess image-classifier errors for safer semi-autonomous robot operation.

Machine-learned labels are checked with voxelized LiDAR intensity, return consistency, and motion data to reduce fog, smoke, and dust false positives.

Ranked object annotations train a neural network to explain roadway decisions, improving scene interpretability and user trust.