Nowadays, drones can be seen in various applications in industry like surveillance and transportation. Industrial drones leverage fully-fledged computer vision techniques, such as object detection based on Deep Learning Neural Networks (DNN), to efficiently perform these objectives. Those techniques come with a high computational effort and are implemented on distributed schemes using ground devices with high performance and power consumption. This limits a drone's operational range since it has to communicate with the ground devices constantly. To alleviate such constraints, an optimized, low-power perception system on the drone is desirable. This work improves a trained DNN architecture to navigate a UAV introduced by the University of Zurich called DroNet. DroNet is computationally expensive and has a high power consumption, making it unsuitable for embedded platforms because of low memory and computational power. In this paper, a ROS-based architecture is first designed to port DroNet on a low-power Jetson Nano board, which conducts the drone's perception and control tasks. Secondly, tuning parameters and various schemes have been carried out to run the inference of the DNN efficiently. To implement the different layers in DNNs, Nvidia's TensorRT SDK is used to compile a high-performance inference engine for the Jetson Nano. Results showed that the Jetson Nano can achieve real-time performance, with 47 frames per second using a Winograd convolution and well-tuned parallelization parameters. The implementation can also achieve a speedup of 2× as compared with the Jetson Nanos ARM CPU while increasing the power consumption by 54%. Finally, the Jetson Nano's usability for drone inference algorithm is shown, achieving real-time response using the DroNet DNN without losing detection accuracy.