Lightweight feature fusion multi-target detection for intelligent assisted driving

In view of the contradiction between the lack of computing resources in vehicle-mounted edge computing terminals and the insufficient accuracy of environmental perception in complex dynamic road conditions, this research proposes a lightweight feature fusion detection model for intelligent assisted driving scenarios. This method uses the improved GhostNetV2 to build a lightweight backbone, and organically integrates the asymptotic feature pyramid network and the efficient multi-scale attention module to solve the semantic loss problem of embedded devices in the feature extraction process. The results revealed that the model achieved an average accuracy of 89.4% and an inference speed of up to 118 frames/second at a very low cost with only 2.6 million parameters and 6.3 gigabytes of floating point operations. Compared with mainstream algorithms, its detection accuracy for long-distance small targets was significantly improved to 71.2%. Moreover, its false detection rate in complex simulation scenarios was effectively controlled. This model achieves deep synergy between detection accuracy and computing efficiency while ensuring millisecond-level response, providing key technical support for the engineering deployment of assisted driving systems on low-computing power platforms.