As the number of car ownership is rising year by year, traffic congestion and traffic accidents are increasingly frequent, and in order to solve such problems, more and more autonomous driving systems based on object detection are proposed, which mainly contain tasks such as road detection, traffic sign detection, and vehicle-pedestrian detection. Among them, detecting key targets such as vehicles and pedestrians is one of the most fundamental and important research branches in autonomous driving scenarios. The existing mainstream research is mainly based on object detection algorithms with predefined anchor frames. Although better results have been achieved, a large number of hyper-parameters need to go to a manual setting, which leads to some limitations in the performance, robustness and generality of the model in complex scenarios. In this paper, an anchor-free based target detection algorithm is applied to investigate the detection of targets such as pedestrians, vehicles and other targets on the road, which has better robustness and generality due to its simple structure and avoidance of hyper-parameters and complex operations associated with anchor frames, and ultimately faster detection rate and higher accuracy. Through experiments on the KITTI dataset, it is demonstrated that the method in this paper has a faster detection rate while the mean confidence score (mCS) can reach 69.1%, and detection performance is significantly better than other previous models.
An Anchor-free Detection Framework for Object Detection in Autonomous Driving Scenarios
2022-12-01
628244 byte
Conference paper
Electronic Resource
English
Free Space Detection for Autonomous Vehicles in Indian Driving Scenarios
Springer Verlag | 2024
|Autonomous Driving Object Detection Platform
Springer Verlag | 2024
|3D Object Detection for Autonomous Driving
Springer Verlag | 2022
|