A Logic-based Compositional Generalisation Approach for Robust Traffic Sign Detection

Zahra Chaghazardi, Saber Fallah, Alireza tamaddoni-nezhad

Keywords: Inductive Logic Programming, Compositional Generalisation, Traffic Sign Detection

TL;DR: Our suggested method for identifying traffic signs is a compositional learning technique that utilises high-level features, which is robust against adversarial attacks and requires only a small amount of training data.

Abstract:

The detection of traffic signs is a fundamental task for Autonomous Vehicles (AVs) to ensure safe and efficient navigation. Although Deep Neural Network (DNN)-based systems play a significant role in developing AV perception systems, they are known to be susceptible to adversarial attacks. This vulnerability is attributed to their dependence on pixel-level features, which can be manipulated to deceive the system and cause misclassification of traffic signs. To address this issue, we propose a logic-based compositional learning approach employing Neural-Symbolic (NS) to detect traffic signs. The proposed methodology decomposes the sign detection task into sub-tasks corresponding to individual sign features, such as shape and text.

We extract these high-level features using OpenCV and Neural Networks (NN) and use an Inductive Logic Programming (ILP) engine to learn and combine the features. This Neural-Symbolic (NS) approach enables our model to capture features and their relationships, making it more reliable to generalise to new and unseen traffic signs. Compositional generalisation is an important challenge in traffic sign detection because traffic signs can appear in a wide range of contexts and configurations. For instance, depending on the country, a "stop" sign could have a different language and configuration. Furthermore, by combining these features, the method is more resilient against adversarial attacks, which makes it better equipped to ensure the safety of all road users.

We evaluated the robustness of our approach by subjecting it to two different adversarial attacks. Our research revealed that the proposed ILP-based technique is able to accurately detect all targeted stop signs, even when exposed to adversarial attacks. Furthermore, this highly efficient methodology demands minimal training data and is fully explainable, which is particularly advantageous in facilitating the debugging of AV systems.