Object detection from roadside edge devices is crucial for enabling efficient vehicle-road collaboration in autonomous driving systems. However, deploying roadside object detection algorithms on resource-constrained edge platforms poses significant challenges due to high computational demands. This research tackles these challenges by prioritizing algorithm lightweighting, computational efficiency, and loss function optimization. A lightweight object detection algorithm is proposed that employs simplification of the network structure and component pruning to minimize parameter count and computation while preserving accuracy. To overcome slow convergence resulting from dataset imbalances, a novel loss function optimization technique is introduced, improving training quality by reducing the impact of low-quality samples. Extensive experimental evaluations validate the effectiveness of the proposed approach, demonstrating its potential to advance perception research in autonomous driving based on vehicle-road collaboration.