Convolutional Neural Network Temporal Sequence Processing

The patent introduces a temporal dimension to the traditional CNN architecture by adding a time axis to the data processing. The input data is organized as a sequence of data items over time, and the CNN processes this extended dimensional structure, allowing temporal dependencies to be captured through convolution operations across the time dimension while maintaining the spatial processing capabilities of traditional CNNs.

The patent segments the processing into distinct functional components: a data generation unit that creates synthetic training data with known time-dependencies, a training unit that trains the CNN on this segmented data, and an evaluation unit that tests the model. This segmentation allows the complex problem of temporal dependency recognition to be solved through structured, modular processing steps.

The patent generates synthetic training data in advance with known ground truth labels and temporal relationships. This preliminary preparation of training data with embedded time-dependencies allows the CNN to learn temporal patterns efficiently during training, reducing the computational burden during actual object detection while maintaining high accuracy.

The patent creates synthetic copies of training scenes with controlled temporal variations and known object trajectories. These copied and modified training examples allow the model to learn robust temporal patterns without requiring extensive real-world data collection and processing, thereby reducing processing time while improving detection accuracy.