Improving Machine Learning-Based Object Detection with Synthetic Data from Ansys AVxcelerate Sensors Simulator

Lidar point cloud-based object detection plays a crucial role in Autonomous Driving. However, obtaining and labelling real-world training data for these systems can be challenging and costly. Synthetic point cloud data offers a cost-effective and flexible alternative for training lidar object detection systems. This study investigates the utilization of synthetic data from Ansys AVX Sensors Simulator to train neural networks and improve lidar object detection in Autonomous Driving. In contrast to many other sensor simulators available on the market, Ansys AVX takes the physics-first approach by accurately modelling the sensor itself (optics, electronics, signal processing) and its environment with the respective high-fidelity optical materials thus enabling the generation of synthetic data that closely resembles real-world scenarios. This study incorporates a mixed training approach, combining synthetic and real-world data, as well as the sequential approach with the pre-training phase using synthetic data followed by fine-tuning on a subset of real-world data. These approaches harness the strengths of both synthetic and real data to simplify and make more accessible the process of training neural networks and improving performance of lidar-based object detection systems. Our experimental results demonstrate the effectiveness of using synthetic data for training, showcasing comparable performance to models trained solely on real-world data. These findings highlight the potential of synthetic data from sensor simulators in advancing Autonomous Driving by addressing the challenges associated with acquiring and labelling real-world data.