Semantic-Aware Bird’s-Eye-View Multi-Sensor Fusion for Enhanced 3D Object Detection in Autonomous Driving

Abstract

The presence of multi- sensor data fusion is everywhere, hence the related study is important. The fusion of data can be found in the real-life activities in several instances. The autonomous driving system that is utilized in the present generation demands a significant comprehension and then a large amount of data to train the system. Experimental data of the imagery and proximity sensors have great importance in the performance of the model. This camera to LiDAR projection is ineffective because the semantic density of the camera is repressed during the process. The current paper tries to refine the traditional point-level fusion methods by placing utmost emphasis on the semantic density. This is made possible through performance optimization as it establishes the hindrances and improves the transformation of the view through the Birds eye View pooling. Moreover, a refinement stage that is confidence-aware is introduced to suppress the occurrences of uncertain features that may occur due to sensor misalignment and noise in an adaptive manner. The LiDAR data is integrated with the camera detections to provide the object tracking with the help of Extended Kalman Filter (EKF). Their detection precision is 0.9546 and the detection recall is 0.9344 with the mAP being assessed as 71.2.