BEVDetNet: Bird's Eye View LiDAR Point Cloud based Real-time 3D Object Detection for Autonomous Driving

3D object detection based on LiDAR point clouds is a crucial module in\nautonomous driving particularly for long range sensing. Most of the research is\nfocused on achieving higher accuracy and these models are not optimized for\ndeployment on embedded systems from the perspective of latency and power\nefficiency. For high speed driving scenarios, latency is a crucial parameter as\nit provides more time to react to dangerous situations. Typically a voxel or\npoint-cloud based 3D convolution approach is utilized for this module. Firstly,\nthey are inefficient on embedded platforms as they are not suitable for\nefficient parallelization. Secondly, they have a variable runtime due to level\nof sparsity of the scene which is against the determinism needed in a safety\nsystem. In this work, we aim to develop a very low latency algorithm with fixed\nruntime. We propose a novel semantic segmentation architecture as a single\nunified model for object center detection using key points, box predictions and\norientation prediction using binned classification in a simpler Bird's Eye View\n(BEV) 2D representation. The proposed architecture can be trivially extended to\ninclude semantic segmentation classes like road without any additional\ncomputation. The proposed model has a latency of 4 ms on the embedded Nvidia\nXavier platform. The model is 5X faster than other top accuracy models with a\nminimal accuracy degradation of 2% in Average Precision at IoU=0.5 on KITTI\ndataset.\n