TinyCenterSpeed: Efficient Center-Based Object Detection for Autonomous Racing

TL;DR

This work presents TinyCenterSpeed, a lightweight adaptation of CenterPoint tailored for 1:10 scaled autonomous racing. By predicting a four-channel heatmap representing position, velocity, and heading from two sequential 64x64 frames, the method achieves real-time perception on CPU with optional TPU offload, and supports multiple opponents without requiring a tracker. Empirical results show up to about 61% improvement over the prior ABD baseline in single-opponent scenarios and robust multi-opponent performance, with quantization and external TPU offload reducing CPU usage by roughly 8x and bringing latency to under 8 ms. The approach offers a modular, open-source solution suitable for constrained autonomy stacks in racing contexts, enabling faster perception without sacrificing interpretability or integration with planning and control.

Abstract

Perception within autonomous driving is nearly synonymous with Neural Networks (NNs). Yet, the domain of autonomous racing is often characterized by scaled, computationally limited robots used for cost-effectiveness and safety. For this reason, opponent detection and tracking systems typically resort to traditional computer vision techniques due to computational constraints. This paper introduces TinyCenterSpeed, a streamlined adaptation of the seminal CenterPoint method, optimized for real-time performance on 1:10 scale autonomous racing platforms. This adaptation is viable even on OBCs powered solely by Central Processing Units (CPUs), as it incorporates the use of an external Tensor Processing Unit (TPU). We demonstrate that, compared to Adaptive Breakpoint Detector (ABD), the current State-of-the-Art (SotA) in scaled autonomous racing, TinyCenterSpeed not only improves detection and velocity estimation by up to 61.38% but also supports multi-opponent detection and estimation. It achieves real-time performance with an inference time of just 7.88 ms on the TPU, significantly reducing CPU utilization 8.3-fold.

Figures (5)

• Figure 1: Schematic depiction of the proposed TinyCenterSpeed architecture to estimate the position, velocity, and heading of opponent autonomous racing cars. The heatmap loss $\mathcal{L}_{hm}$ is inspired from CenterPoint, but the architecture is optimized for computationally constrained such as encountered in 1:10 scaled autonomous racing.
• Figure 2: This figure exemplarily shows the architecture of TinyCenterSpeed. The figure accurately depicts the model dimensions and shows the convolutional and deconvolutional layers as well as the residual connections used in the architecture. Note that the images of the robots are for visualization purposes only and are not used in the actual model.
• Figure 3: This figure shows an exemplary multi-channel output heatmap $\hat{\textbf{H}}$ consisting of channels for the position, velocity in both x- and y-direction, and the estimated yaw angle.
• Figure 4: The physical ego car behind the opponent car is shown here in \ref{['subfig:physical_cars']}. The preprocessed input and the corresponding positional ground truth are instead shown in \ref{['subfig:data']}.
• Figure 5: The ForzaETH racecar with the sensor, the and the external Google Coral connected via USB. This allows offloading of the inference from the main to the when the is INT8 quantized.