Abstract: Addressing the deficiencies in reliability and real-time performance of UAV perception-decision-making chains in complex environments, this study integrated and accelerated optimized versions of an improved real-time object detection algorithm (YOLOv8s) and an enhanced A^* dynamic path planning algorithm on the Jetson Orin NX edge computing platform. Within the object detection module, the YOLOv8s algorithm was improved by modifying its neck structure and introducing GSConv-based lightweight convolution along with the SlimNeck feature optimization module. Combined with TensorRT INT8 quantization acceleration, this achieved a detection speed of 57 FPS while maintaining 72% accuracy for small object detection. For the path planning module, the A^* dynamic path planning algorithm was enhanced by integrating obstacle confidence prediction and kinematic models, along with restructuring the path cost function. Utilizing CUDA parallel acceleration, the planning computational complexity was reduced by 22%, average latency decreased by 80.6%, and collision probability lowered by 86.3%. The overall system end-to-end latency was reduced to below 400 ms. UAV field tests conducted in real-world environments validated the effectiveness and practicality of this approach in enhancing UAV autonomy within complex scenarios.