Autonomous driving can enhance road safety and make transport more convenient. In recent years, numerous research works have focused on autonomous driving. However, an underlying problem with autonomous cars is that the Deep Learning-based object detection techniques sometimes get caught, giving false negatives.
A new study has been published wherein researchers have suggested various techniques that would help avoid false negatives. The approach could be a game-changer for the Autonomous Vehicles Market. Furthermore, the proposed idea can also facilitate interpretable probabilistic predictions. There is no requirement of re-training the network, so the technique becomes appealing due to its practicality.
False positives refer to a situation where any obstacle or object is not there but gets identified by the system. If the problem of erratic breaking occurs, it may affect the person's safety or the vehicle's overall safety.
The centerpiece of autonomous driving is considered to be Object Detention. Usually, today's DL methods use a single value acquired through the SG (Sigmoid Function) or an SM (Softmax Function). These functions work by exporting the detection confidences like normalized scores. However, they do not consider overconfidence or uncertainties while giving predictions. Thus, the suggestions sometimes bring out overconfident predictions that are just false negatives.
In the present study, researchers have proposed a unique probabilistic layer that successfully evades the false predictions made by Softmax or Sigmoid. The probabilistic methodology is validated with the help of multi-sensory 2D and 3D object detection. This is done through ReV (Reflectance-View), RG images and RaV (Range-View) maps modalities.
The research suggests that conventional prediction layers could produce bad decision-making within deep object detention networks. So, they provided an alternate way to acquire correct probabilistic inference through MAP (Maximum a-Posteriori) and ML (Maximum Likelihood). The technique is also validated with the help of 2D-KITTI objection detection via SECOND (Lidar-based detector) and the YOLO V4.
The interesting part of the research is that the team has demonstrated a technique to reduce overconfidence of false positives within approaches. Further, this was achieved without disturbing the performance of true positives. The study can benefit the autonomous vehicle sector to make it more successful.