Optimizing Lightweight Object Detection Models for Autonomous Driving: A Comparative Study of Model Compression, Real-Time Performance, and Transfer Learning for Resource-Constrained Devices

This paper discusses strategies for optimizing lightweight object detection models in autonomous driving, with regard to performance improvements on resource-constrained embedded platforms, such as the Nvidia Jetson and Raspberry Pi. The key optimization techniques involved are channel pruning and quantization, reducing model size and computational complexity that improve inference speed and efficiency. It also discusses the improvement of real-time detection speed, including lightweight architectures and pipeline optimizations to meet the stringent frame rate requirements of autonomous vehicles. It also discusses methods for improving detection accuracy in complex environments, such as urban streets and adverse weather conditions. The paper emphasizes the importance of balancing efficiency, accuracy, and speed to ensure the feasibility and safety of object detection in autonomous driving systems.