Autonomous driving can improve road safety and make transport more efficient. A lot of research has been focused on autonomous driving in recent years. Deep Learning based object detection techniques sometimes give false negatives. G. Melotti, W. Lu, D. Zhao, A. Asvadi, N. Gon calves, and C. Premebida have discussed ways of resolving this issue in their research paper titled “Probabilistic Approach for Road-Users Detection” which forms the basis of the following text.
Autonomous driving: a car undergoing road testing. Image credit: Dllu via Wikimedia, CC-BY-SA-4.0
Why this Research is Important for the Autonomous Driving?
False positives mean situations where an object or obstacle is not there but was detected by a system. Erratic braking in such a situation affects the person’s safety and the vehicle’s overall state. The researchers have proposed a technique that aims to avoid these false positives, thereby being a game-changer for adopting autonomous vehicles. Also, the proposed approach enables interpretable probabilistic predictions. Without re-training the network, it makes the technique practical.
Description of the Proposed Algorithm
Object Detection is the centerpiece of autonomous driving. Generally, modern DL methods use Softmax function (SM) or a single value obtained from the Sigmoid function (SG). These functions export the detection confidence as the normalized scores without considering the overconfidence or uncertainties in the predictions. Hence, this prediction could sometimes produce overconfident predictions for false positives.
Image credit: arXiv:2112.01360 [cs.CV]
YOLO V4 framework is used for object detection. The above image demonstrates YOLO V4 representation with Logits and Sigmoid (SG) layers, Maximum Likelihood (ML) and Maximum aPosterior (MAP) functions. After training, the predicted values from the Sigmoid Layer were replaced by the scores from ML and MAP functions. We should note that the YOLOV4 was not trained or re-trained with the ML/MAP functions.
The researchers have proposed a novel probabilistic layer that avoids the traditional Sigmoid or Softmax prediction layer in this research. The proposed probabilistic methodology is validated through multi-sensory 2D and 3D object detection using RGB images, range-view (RaV), and reflectance-view (ReV) maps modalities.
The research showed that traditional prediction layers could induce erroneous decision-making in deep object detection networks. The researchers have proposed an efficient way to obtain proper probabilistic inference via Maximum Likelihood (ML) and Maximum a-Posteriori (MAP) formulations. This technique is validated on the 2D-KITTI objection detection through the YOLO V4 and SECOND (Lidar-based detector)
The researchers have demonstrated that the proposed technique reduces overconfidence in false positives without degrading the performance of the true positives. In the words of the researchers,
This paper proposes a formulation (called ML/MAP layers) to reduce the overconfidence of detected false positive objects without degrading the classification scores of true positives i.e., the ML/MAP layers are be able to reduce confidence in incorrect predictions. The formulation takes into account a probabilistic inference through two models, one being non-parametric (normalized histogram) and the other is parametric (Gaussian density to model the priors for the MAP). As a way to present the efficiency of the proposed probabilistic inference approach, this work considered different modalities, as RGB imagens, RaV, and ReV maps, as well as 3D point clouds data i.e., datasets with different characteristics. In the case of RGB images, the characteristics are obtained directly from the camera, while RaV and ReV maps are obtained from depth (range-view) and intensity (reflectance-view) data, respectively. The results achieved by the proposed approach are very satisfactory, specially for the minority category ‘cyclists’ (for YOLOV4), and ‘pedestrian’ case (for SECOND), as evidenced by the performance measures (Pr-Rc curves and AUC). Finally, a key advantage of the proposed approach is that there is no need to perform a new network training, that is, the approach has been applied in already trained networks
Source: G. Melotti, W. Lu, D. Zhao, A. Asvadi, N. Gon¸calves and C. Premebida, “Probabilistic Approach for Road-Users Detection”