Hardware and firmware tuning for point cloud object detection in embedded systems

Object detection in point clouds is a central aspect of many robotics applications such as autonomous driving. Real-time technologies require very high speed devices and demand low complexity algorithms, at the expense of accuracy. In this study we consider the trade-off between inference time and accuracy of an object detection model. In particular, our purpose is to answer the following question: how much can we reduce the inference time of said algorithm maintaining a sufficient accuracy and keep satisfactory performance? To solve this problem we exploit the PointPillars algorithm, an encoder that utilizes PointNets to learn a representation of point clouds organized in vertical columns (pillars), which outperforms many other methods with respect to both speed and accuracy by a large margin. Tuning is applied to some parameters of this model, such as the number of filters of the feature encoder and the number of layers of the backbone, without changing its global structure. Through training and testing performed with the KITTI benchmark, we obtain the trends of the accuracy versus inference time relations along the applied modifications. Studying these tendency functions, we extrapolate the optimal solution which lowers the validation time without significantly reducing the performance. This solution consists in the reduction in the amount of layers in the backbone, and in the number of up-sample filters at its output. On this basis, after building the environment to work on Nvidia Jetson Nano, an embedded system that contains a GPU for high-performance computing tasks, future work could concentrate on applying the improved model to this machine, with the aim of analysing its power efficiency and performance.