Autonomous vehicles play a crucial role in logistics, agriculture, and warehousing, requiring precise object detection and recognition for safe navigation in confined spaces. Traditional 2D sensor-based methods and simple LiDAR applications often struggle with depth perception and classification accuracy, limiting real-time decision-making. This study addresses these challenges by developing a custom LiDAR-based dataset for object recognition within the Voxel-RCNN framework, focusing on six object categories to enhance recognition accuracy. The Voxel-RCNN model was trained on this custom dataset without architectural modifications, assessing its generalization to non-standard data and performance in constrained environments. The training process demonstrated stable convergence, with loss decreasing from 6.09 to 2.37 after 600 epochs. Quantitative evaluations using BEV and 3D Average Precision (AP) revealed strong performance in detecting structured objects like cars (68.14% BEV AP, 55.83% 3D AP in Easy cases) but significant challenges with occluded and irregularly shaped objects such as trees and cyclists. Despite these challenges, the study highlights the potential of Voxel-RCNN for 3D object recognition in autonomous navigation. Future improvements include dataset augmentation, multi-scale feature fusion, and advanced voxelization techniques to enhance recognition performance. These findings contribute to the advancement of LiDAR-based perception systems, supporting safer and more intelligent autonomous vehicle operations.