Analysis of Model Optimization Strategies for a Low-Resolution Camera-Lidar Fusion Based Road Detection Network

Road detection is one of the primary tasks for autonomous vehicles. However, it can be challenging to detect the road or the drivable region with only color cameras on the unstructured road. In this study, a low-cost and low-resolution Camera-Lidar fusion based deep segmentation network is proposed to detect the front road region. The fusion network can capture both the color features from the image and spatial features from the point cloud, which can be significantly effective in the unstructured road condition. A deep segmentation convolutional neural network is designed to process both the RGB image and the calibrated Lidar point cloud. The Lidar coordinate system was first transformed into the camera coordinate, and the 3D point cloud information is transformed to the 2D color image, which maintained the depth features and textures of the road. A low-cost, low-resolution solution is proposed by rescaling the original high-resolution images into a low-resolution format to increase the real-time inference speed. A cross fusion segmentation network is trained to process the two different inputs simultaneously. To evaluate the efficient model optimization methodologies, several different criteria and learning rate adjustment methods are evaluated. The models are trained and tested with the KITTI public dataset. Results indicate that low-cost the cross-fusion network can provide a reasonable road detection with an exponential learning rate adjustment.