Online Sensor Selection for Object Detection via Bayesian Risk Minimization
Abstract
In today's world, object detection systems increasingly rely on multiple sensors, but processing data from all of them is computationally, energy, and bandwidth intensive. Situations such as battery-saving mode, limited network bandwidth, processor cooling requirements, or handling other tasks simultaneously make it difficult to use all sensors at once. This motivates the need for methods that dynamically select sensors, achieving high detection performance while conserving resources. We propose a decision-theoretic approach for object detection that selects the optimal sensor based on incoming evidence, rather than using all sensors simultaneously. Using Bayesian principles, the system chooses the sensor that minimizes the expected risk of using that sensor, combining the cost of computation and detection errors given the evidence. This dynamic selection balances high detection performance with limited resource usage, producing an output expected to have minimal risk. To demonstrate our sensor selection method, we applied it to choosing between analyzing a low-resolution (LR) or high-resolution (HR) image for car detection. In this setup, the system uses the output of the car detector on the low-resolution image as evidence to decide whether the corresponding high-resolution image should be analyzed for a given region. Our approach, which selects online between high- and low-resolution sensors, increased recall from 0.66 to 0.71 at 0.75 precision compared to using only the medium-resolution sensor on the KITTI dataset. It also reduced processed pixels by 34% and outperformed the medium-resolution pipeline in both computation and detection performance. Overall, our method outperforms low-, medium-, and high-resolution pipelines in terms of total risk.