Robust Object Detection Using Depth from Monocular Camera for Feature Common Representation
Hyunmin Kong, Jitae Shin · 2024
In autonomous driving scenarios, harsh conditions such as bad weather, sensor contamination and obstructed visibility are common. These factors directly contribute to the degradation of detection performance, posing serious safety threats. Recent studies have attempted to address these challenges by improving the robustness of detection systems through the integration of 3D spatial information, such as depth [1]. However, commonly used sensors for acquiring 3D spatial information, such as LIDAR, LADAR and stereo cameras, often suffer from drawbacks such as high cost and sensor limitations. To address these issues, this paper proposes a solution using relatively inexpensive RGB cameras. This technique uses a monocular camera to estimate depth based on RGB images, eliminating the need to synchronize RGB and depth images. The effectiveness of the proposed method was evaluated using the KITTI 2D dataset, which experimentally validated improvements in all metrics including mAPSO, mAPSO-9S, mAP7S, mAPs, mAPm and mAPI in noisy data. These results demonstrate that the method proposed in this paper can operate robustly in harsh environmental conditions such as bad weather, even without the use of expensive sensors