Strategic Improvements of SqueezeSegV2 for Road-Scene Semantic Segmentation Using 3D LiDAR Point Cloud

Abstract

Semantic segmentation of LiDAR point clouds for road-scene analysis in autonomous vehicles and driver assistance systems is a challenging task due to the confusion of categories and the sparse distribution of point clouds, thus leading low performance. In this paper, we propose two important improvements to SqueezeSegV2, a deep encoder-decoder neural network, to improve the overall performance of semantic segmentation. The first improvement is the adaptive Fire module, which can be configured to be lightweight or accurate, depending on the service and application requirements. The second one is the steady Fire Deconvolution module, which boosts the accuracy of the segmentation mask reconstruction. Remarkably, both modules are improved by apply manipulating the combination of symmetric and asymmetric grouped convolution with dilation rate to enhance the contextual learning efficiency of the deep model. We evaluate our proposed methods on the Panda dataset and show that they achieve better segmentation accuracy than the original SqueezeSegV2 model by mean accuracy and mean IoU, while also reducing the number of trainable parameters by around .