Overcoming forgetting in artificial neural network (ANN) for long range object detection

• Main Author: Wu, Jiayu
• Other Authors: Soong Boon Hee
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2025
• Subjects: Engineering; Long-range detection; Deep neural networks
• Online Access: https://hdl.handle.net/10356/181928
• Institution: Nanyang Technological University

Self-driving cars represent a significant advancement in transportation, but face many challenges in practical applications, such as reliable scene sensing and decision-making under variable environments and data distribution. LiDAR and Radar are key sensors in self-driving cars. However, problems like bad weather can pose serious challenges for object detection. Moreover, deep neural network-based object detectors perform well in autonomous driving scene perception, but in real-world applications, model performance may be degraded due to variations in the operational domain and data distribution. The model can also suffer from catastrophic forgetting during the domain incremental learning process. This occurs when a model's performance in previously learned domains deteriorates after adapting to new data. In this project, we investigated the well-established RADIATE dataset to examine the impact of catastrophic forgetting in adverse weather conditions for both LiDAR and Radar sensors. We conduct experiments using RADAR and LiDAR detection networks to identify the most effective training strategy for reducing forgetting. Our experiments reveal that by pre-training the detection framework and freezing 15 layers at the fine-tuning stage leads to a reduction of 4.68% of in forgetting. Corresponding to these findings, we conclude that catastrophic forgetting could be caused by changes in low-level features, and we can effectively reduce this issue by finding the most appropriate number of freezing layers for fine-tuning to enable the model to learn the generalized features of data. Furthermore, we explored the factors contributing to reduced detection accuracy in adverse weather conditions. Our findings indicate that long-range object detection (at distances of 60 meters and beyond) presents the most significant challenge. The primary obstacle is the sparseness of the LiDAR point cloud at greater distances, which impairs the ability to detect long-range objects, especially in adverse weather. Additionally, annotating LiDAR data under these conditions becomes increasingly difficult as the target point cloud becomes sparser. To address this issue, we fine-tuned the detection module using both real and simulated long distance data, thereby enhancing the accuracy of long-range detection. This approach also helps mitigate the phenomenon of catastrophic forgetting more effectively. With freezing 15 layers and simulations at the fine-tuning stage, it can reach a reduction of 6.27% of in forgetting.