Object detection algorithm using machine learning & sensor fusion techniques for autonomous vehicle applications

Autonomous vehicles perception system is one of the most important modules for any autonomous driving applications in which autonomous vehicles are equipped with sensors that enables the vehicle to generate massive amount of data. One of the biggest challenges is transforming these sensory data into semantic data such as accurately identifying all the different road users. Failing to detect and avoid objects can lead to tremendous repercussions such safety related incidents.

This has seen rapid development especially with the advancement and assistance of machine and deep learning techniques. However, objects in large-scale and sparse point clouds such as LIDAR data are hard to detect accurately especially in terms of reducing the searching area when detecting objects in point clouds is a key challenge. Therefore, to leverage 2D images for solving such challenge, we proposed a 2D-driven 3D object detection framework using point cloud datasets. However, in adverse weathers, LIDAR tends to collect many noisy points from rain or snow, which may disturb the results of object detection, resulting in false detection. To enhance the robustness of the detector, we improve existing LIDAR only 3D object detector from two aspects under different weather conditions.

This novel model was validated on the CARLA-AUTOWARE simulator and KITTI with a val set of 88.77%, 78.13%, and 75.29 % 3D car AP for easy, moderate, and hard difficulties, respectively.