Latent Heatmap Training for Occluded Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Deep neural networks (DNNs) used in object detection for autonomous driving struggle with challenging data samples, such as occluded objects, leading to missed detections and hazardous events.

Innovation Solution

A method for training neural networks for object detection that involves generating latent heatmaps from neural network activations, calculating a heatmap loss, and backpropagating this loss for training. This approach aims to produce meaningful activation maps with minimal noise, improving the detection of occluded objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If standard object detection methods are used, then detection speed is maintained, but detection precision for occluded objects deteriorates

Engineering Contradiction:
Improvedetection precisionVSAvoidreliability for occluded objects
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces latent heatmaps as an intermediary representation between the input image and final detection output. These heatmaps serve as a mediator that highlights potential object locations and guides the detection process, improving both precision and reliability for occluded objects without sacrificing detection speed

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The detection process is segmented into multiple components: backbone feature extraction, latent heatmap generation, and detection head processing. This segmentation allows each component to specialize - the backbone handles general feature extraction while the latent heatmap specifically addresses occluded object detection, improving overall reliability

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If more complex detection algorithms are applied, then detection precision improves, but computational complexity increases

Engineering Contradiction:
Improvedetection precisionVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges the latent heatmap generation with the existing backbone feature extraction process. By reusing features from the backbone and integrating heatmap generation within the same computational framework, the system improves detection precision without proportionally increasing computational complexity

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

Instead of completely redesigning the detection algorithm, the patent applies partial action by adding latent heatmap generation to specific parts of the existing pipeline. This selective enhancement improves precision for critical cases (occluded objects) while maintaining efficiency for standard detection scenarios

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4553788A1Object detection using a trained neural network
Publication Date: 2025.05.14 VOLKSWAGEN AG
  • EP4553788A1 patent drawingFigure 1
  • EP4553788A1 patent drawingFigure 2~3
  • EP4553788A1 patent drawingFigure 4

AI summary

The present invention is related to a method, a computer program, and an apparatus for training a neural network for object detection, and to a respective trained neural network. The invention is further related to a method, a computer program, and an apparatus for object detection with a trained neural network, and to a vehicle equipped with a perception system, which makes use of such a method or apparatus. In a first step, an input image is received (10), which is then processed (11) with the neural network. Latent activations for the image are extracted (12) from one or more layers of the neural network and a latent heatmap is generated (13) from these latent activations. If not yet available, a ground truth mask for the input image may be generated (14) by converting bounding boxes representing object instances into a binary mask. A heatmap loss between the latent heatmap and the ground truth mask is then determined (15). At least the heatmap loss is backpropagated (16) for training the neural network.