VisiDriveNet: A Deep Learning Model for Enhanced Autonomous Navigation in Urban Environments

Autonomous vehicle technology has progressed significantly; however, navigating complex urban environments remains challenging. Current systems using single-sensor modalities such as cameras, LIDAR, or RADAR struggle with accurately perceiving dynamic urban landscapes, handling obstacles, varying lighting conditions, and unpredictable pedestrian and vehicle movements. To address these issues, the proposed work introduces VisiDriveNet, an architecture that uniquely integrates LIDAR and RADAR data using an advanced FusionNet-CNN-LSTM deep learning model. This model combines CNNs' spatial feature extraction of CNNs with LSTMs' temporal sequence learning of LSTMs to effectively process fused sensor data. Optimized for real-time processing, VisiDriveNet enhances the safety and reliability of autonomous navigation. Developed and evaluated using the KITTI Vision Benchmark dataset, which includes 7481 training and 7518 test images with the corresponding LIDAR and RADAR data, VisiDriveNet achieved superior results. Performance metrics show precision ranging from 91.2% to 88.5% and IoU from 86.4% to 83.7% across various urban scenarios, significantly outperforming the existing systems. Future research will focus on improving performance under variable conditions and further optimizing real-time processing to reduce latency.