Algorithm Evolution and Technical Challenges in Autonomous Driving Object Detection

Abstract

The rapid evolution of autonomous vehicle technologies has positioned object detection systems as pivotal components for reliable environmental perception. This review systematically examines three critical dimensions: architectural advancements from traditional CNN-based models to Transformer architectures, strategies for mitigating environmental interference, and practical implementation challenges in edge computing. Through comprehensive analysis of 35 peer-reviewed studies (2018–2023), Transformer-based models demonstrate a 12.7% improvement in mean average precision (mAP) over single-stage detectors in complex urban scenarios, albeit with a 43% increase in computational latency. A significant dataset bias is identified, with nighttime samples constituting less than 4.7% of major benchmarks, directly correlating with 22–35% performance degradation under low-light conditions. To address these limitations, a hybrid quantization-distillation framework is proposed, integrating neural architecture search-based channel pruning, adaptive mixed-precision quantization, and attention-guided knowledge transfer. Experimental validation on NVIDIA Jetson AGX Xavier platforms achieves 94.6% model compression efficiency while retaining 89.3% of baseline accuracy. These findings establish guidelines for developing next-generation perception systems that balance computational efficiency (≤50ms latency) with detection reliability (≥92% mAP) in dynamic environments.