Abstract:Robust 3D object detection is a core challenge for autonomous mobile systems in field robotics. To tackle this issue, many researchers have demonstrated improvements in 3D object detection performance in datasets. However, real-world urban scenarios with unstructured and dynamic situations can still lead to numerous false positives, posing a challenge for robust 3D object detection models. This paper presents a post-processing algorithm that dynamically adjusts object detection thresholds based on the distance from the ego-vehicle. 3D object detection models usually perform well in detecting nearby objects but may exhibit suboptimal performance for distant ones. While conventional perception algorithms typically employ a single threshold in post-processing, the proposed algorithm addresses this issue by employing adaptive thresholds based on the distance from the ego-vehicle, minimizing false negatives and reducing false positives in urban scenarios. The results show performance enhancements in 3D object detection models across a range of scenarios, not only in dynamic urban road conditions but also in scenarios involving adverse weather conditions.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the challenges faced by LiDAR - based 3D object detection models in practical applications in dynamic urban environments, especially how to reduce false positives and false negatives. Specifically, existing 3D object detection models perform well when dealing with near - distance objects, but their performance degrades when detecting far - distance objects, resulting in a large number of false positives in complex urban scenes, which may cause sudden stops or accidents in autonomous driving systems. Therefore, the paper proposes a distance - based adaptive threshold algorithm to dynamically adjust the detection threshold and improve the robustness and accuracy of 3D object detection.
### Main contributions of the paper:
1. **Distance - based adaptive threshold**: By introducing a distance - based adaptive threshold in the post - processing stage of the 3D object detection model, false positives for near - distance objects and false negatives for far - distance objects are significantly reduced, and the robustness of the model in dynamic urban environments is improved.
2. **Simplified integration process**: The proposed algorithm can be seamlessly integrated into existing 3D object detection frameworks without additional training or unnecessary complexity.
3. **Multi - framework verification**: Experimental results show that the algorithm can effectively improve the detection performance in multiple 3D object detection frameworks, especially in terms of the balance between mean average precision (mAP), recall, and precision.
### Specific methods:
- **Adaptive threshold formula**:
\[
\text{Score Threshold} =
\begin{cases}
\alpha d^2+\beta d + \gamma & \text{if } 0 < d\leq\delta \\
k & \text{if } d > \delta
\end{cases}
\]
where \(d\) is the distance between the object and the host vehicle (in meters), \(\alpha\), \(\beta\) and \(\gamma\) determine the shape of the quadratic curve of the confidence score as a function of distance, \(\delta\) is the maximum distance to which the algorithm applies, and \(k\) is the constant value of the quadratic curve at distance \(\delta\).
- **Parameter optimization**:
By analyzing the mean and standard deviation of the confidence scores in different distance intervals, the parameters of the adaptive threshold formula are determined as \(\alpha=- 0.00002\), \(\beta = - 0.0061\), \(\gamma=0.6828\) and \(k = 0.6\).
### Experimental results:
- **Quantitative evaluation**: Experiments were carried out on the Kitti 3D object detection dataset and a custom - made urban road dataset. The results show that the adaptive threshold algorithm can significantly reduce false positives and false negatives in multiple 3D object detection models (such as PointPillars, SECOND, PointRCNN and PV - RCNN), and improve the overall detection performance.
- **Qualitative evaluation**: In tests on actual urban roads and in adverse weather conditions (such as fog and rain), the adaptive threshold algorithm can identify vehicles more accurately, reduce false positives caused by point - cloud noise, and ensure the safety and reliability of autonomous driving systems.
In conclusion, by introducing a distance - based adaptive threshold algorithm, this paper solves the problems of false positives and false negatives faced by 3D object detection in dynamic urban environments, and improves the robustness and safety of autonomous driving systems.