An Acceleration Inference Implementation of BEVFusion with MQBench on Xavier

The integration of sensor technology and deep learning has made multi-sensor fusion algorithms an essential component of autonomous driving perception. Among these algorithms, BEVFusion performs well in 3D perception tasks. However, its high computing cost presents challenges for deployment and application on the hardware. In this work, we propose a way that can trade off efficiency and performance to compress and deploy it on Jetson Xavier NX. We first decouple the model into three parts and then use MQBench to quantize the Fusion and Decoder of BEV in various ways. After that, we accelerate it on the hardware with the parameters of those schemes by utilizing TensorRt.By analyzing the mAP and inference time of the model in different quantization factors, we obtain the optimal schemes when using FP32 as scale, per-channel as granularity and EMAMinMax as calibration algorithm, whose accuracy only experiences a marginal drop of 0.12% while the execution speed witnesses a remarkable increase of 98.3% compared to the original model. The significant improvement makes the application of BEVFusion feasible for real-time autonomous driving perception.