Comparative Analysis of Effective AI based 3D Multi-Object Detection and Tracking Methods for Autonomous Driving

Abstract

Object detection is a crucial task for autonomous driving, and different autonomous vehicles have varying perceptions. The advancements of object detection paved the way for 3D object detection, which is considered to be the central component of perception systems that predict obstacles, vehicles, pedestrians and other key features of the environmental backgorund. Generally, various sensors and cameras are used in autonomous driving producing accurate perediction of objects. Several algorithms have been employed in object detection, but they have not produced effective outcomes.

Thus, the present study implements HDL-MODT (hybrid deep learning based multi-object detection and tracking) using a sensor fusion approach. It uses solid-state LiDAR, pseudo-LiDAR and an RGB camera to capture objects and provide effective tracking abilities. Initially, the pre-processing methods involved noise removal using an A-Fuzzy (adaptive fuzzy) filter. Contrast enhancement is then performed using the MSO (moth swarm optimization) algorithm, and feature segmentation is done by LGAN (lightweight general adversial networks), where both channel and position attention mechanisms provide precise segmentation. The YOLOv4 approach is deployed for detection of objects such as ground, vehicles, pedstrians and obstacles. Finally, the tracking of objects is performed using IUKF (improved unscented Kalman filter). The simulation of the proposed method is demonstrated by using MATLAB R2020 simulation tool; the performance of the proposed method is also predicted by comparing the results with exisitng algorithms.