A Two-Stage Model Compression Framework for Object Detection in Autonomous Driving Scenarios

Recently, object detection, as a critical technology to improve the autonomous perception capabilities of optical sensors in autonomous driving systems (ADSs), has become a primary research focus in the field of ADS perception. However, the practical implementation of these networks can be challenging due to their scale and complexity, making it difficult to implement them directly on devices with limited resources. To address this issue, a universal two-stage model compression approach has been implemented. During the initial phase, ShuffDet (SD) is introduced as a lightweight network architecture to reduce the structural parameters within the network effectively. During the second phase, probability distribution distillation (PDD) techniques are applied to the network post-lightweighting to mitigate the impact of structural lightening on network precision. The algorithm was tested using two public datasets, BDD100K and KITTI. The experimental outcomes demonstrate that this method enhances precision while substantially lowering the model’s complexity. To demonstrate its universality, we replaced the base network with YOLOX, which produced satisfactory results. To determine the effectiveness of the method in real-world deployment settings, we deployed the model on an NVIDIA Jetson Nano chip. The experimental outcomes confirmed the effectiveness of our proposed approach, achieving real-time detection standards. When compared to alternative lightweighting techniques, this method is more advantageous for deployment in ADSs.