This methodology shows the on-device CNN inference in autonomous UAV navigation using a low-power multi-core architecture. The system has employed the Convolutional Layer Feature Extraction with pre-trained CNN models, so that fast extraction of features of the input images can be run off from the system. The MobileNetV2, lightweight CNN with minimal engagement overhead is used, but guarantees high performance in terms of classification performance. Thus, it is applicable in resource-restricted devices. It uses tensorflow lite architecture customised to low-power, embedded systems, and delivers cost-effective inference and energy-efficient processes. The suggested method contributes to considerably lower power dissipation, inference latency with a high accuracy in classification. It can be used to provide real-time obstacle detection, dynamics path planning, and navigation of UAV based systems and this makes it appropriate in longer operating missions of autonomous UAVs. This has been demonstrated in the results obtained indicating that the system provides an effective solution to scalable, efficient implementation of CNN based models in UAVs, thus leading to more intelligent and energy-efficient autonomous systems in many applications.