Obstacle Detection and Avoidance in Self-Driving Carsusing LIDAR and Vision-Based Algorithms

This is the authors' abstract. We don't add key points for this paper.

Existing algorithms do not generalize well to the Light Detection and Ranging (LIDAR) and camera modality when trained to detect cars with a vision-based algorithm, or vice versa when LIDAR is used as a separate sensor. These systems would fail in severe conditions such as low lighting and inclement weather. Itpresents a system that combines Lidar and vision-based data utilizing sensor fusion to enable Lidar's depth perception capabilities and vision-based algorithms for encompassing object recognition tasks. Such technology increases obstacle detection, object classification, and real-time decision making to provide a safer autonomous navigation system. It combines Kalman Filtering, convolutional neural network (CNN) for object detection, density-based spatial clustering of applications with noise (DBSCAN) for LIDAR data clustering, support vector machine (SVM) for classification, and predictive path planning. The results demonstrate the following improvements: 96.5% detection accuracy, 1.5% false positive rate, and 0.2% collision rate. It is more robust and outperforms existing systems in real-world situations such as bad illumination, severe weather, and complex traffic scenarios.