Recurrent Convolutional Feature Pyramid Network for Radar Trajectory Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional object tracking methods using radar data face challenges in distinguishing targets from background and accurately estimating size and rotation, especially for stationary targets due to noisy and sparse data.
Innovation Solution
A computer-implemented method that combines a recurrent convolutional one-stage feature pyramid network for joint object detection and motion forecasting, utilizing radar data to predict trajectories by propagating information through multiple network layers and using a pyramidal structure to generate intermediate data for efficient and reliable tracking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional clustering and recursive filtering methods are used for multi-object tracking on radar data, then the tracking problem can be divided into motion prediction and data association, but the noisy and sparse radar data causes issues in distinguishing targets from background and correctly estimating size and rotation
Solution Approach 1:
The patent combines object detection and motion forecasting into a single joint network that processes radar data simultaneously for both tasks. This merging eliminates the need for separate clustering and filtering steps, directly addressing the noise and sparsity issues by learning features end-to-end rather than through multiple sequential processing stages.
Solution Approach 2:
The patent introduces a feature pyramid structure that processes radar data at multiple scales and temporal resolutions. By analyzing data across different dimensional scales (spatial and temporal), the network can distinguish targets from background more effectively and estimate target properties more accurately, overcoming the limitations of conventional single-scale methods.
2Productivity
If conventional methods are used for stationary targets, then processing can be simpler, but accurate estimation of size and rotation becomes particularly difficult due to noise and sparsity
Solution Approach 1:
The network performs preliminary feature extraction and temporal pattern recognition on radar data before final target detection and parameter estimation. By preparing and processing features in advance through the recurrent network and feature pyramid, the system maintains high reliability for stationary targets without sacrificing processing efficiency, as the heavy lifting is done during the initial joint detection-forecasting pass.
3Measurement precision
If joint detection and motion forecasting is performed on radar data using a recurrent convolutional feature pyramid network, then object detection and trajectory prediction are improved, but the network architecture becomes more complex
Solution Approach 1:
The patent segments the neural network into distinct functional modules: a recurrent backbone for temporal feature extraction, a feature pyramid for multi-scale processing, and separate detection/forecasting heads. This segmentation allows each component to specialize in specific tasks, improving overall detection precision while making the complex architecture more manageable and trainable through modular design.
Data Source
AI summary
A computer implemented method for predicting a trajectory of an object comprises the following steps carried out by computer hardware components: acquiring radar data of the object; determining first intermediate data based on the radar data based on a residual backbone using a recurrent component; determining second intermediate data based on the first intermediate data using a feature pyramid; and predicting the trajectory of the object based on the second intermediate data.