Computer vision is significantly contributing to and setting up new benchmarks in image and video analysis. Camera sensors are analogous to human vision systems capturing relevant information inside cabins as well as external environs of an autonomous vehicle and reckoning the data collected with a brainy intelligence discharged by appropriate deep learning algorithms operating an embedded edge platform with sufficient configurations. The existing algorithms have conquered the ace accuracy levels but are yet struggling to achieve the desired inference. To cope with these challenges diverse ingredients such as PyTorch framework, standard for interoperability of deep learning models-Open Neural Network Exchange (ONNX) and model optimizer-TensorRT are cooked with tempering of model compression for high performance inference recipe. Our work manifests inference performance for FaceNet face recognition model coalescing with unauthorized access for in-cabin surveillance in autonomous vehicles. The inference significantly improves by almost double with TensorRT FP16 precision mode and by quadruple for int8 precision mode. There is a reduction in accuracy by 0.5%, however improving the inference performance by 34% when migrating to TensorRT FP16 mode from TensorRT FP32 mode.