Acceleration of Image Classification and Object Tracking by the Intel Neural Compute Stick 2 with Power Efficiency Evaluation on Raspberry Pi 4B

This work investigates the efficiency and power consumption of using the Intel^® (Santa Clara, CA, USA) Neural Compute Stick 2 (NCS2) on the Raspberry Pi 4B platform to accelerate image classification and object tracking. The motivation behind this study is to enable the real-time operation of complex neural networks in embedded systems, potentially reducing the cost of deep learning neural network deployment and expanding industrial applications. This study also supplements the OpenVINO™ 2022.3.2 documentation by recording the application of the Raspberry Pi 4B combined with the NCS2 in the latest European software repositories. Supported by OpenVINO™ 2022.3.2 and the Deep SORT algorithm, this study consists of two distinct tests: image recognition and real-time object tracking. A single model is used for image recognition, while two models are deployed for object tracking. These test cases evaluate the performance of the execution hardware by varying the different number of models in different application scenarios and evaluating the impact of NCS2 acceleration under various conditions. The results indicate that, for the specific models used in this experiment, the NCS2 increases image recognition performance by approximately 400% and real-time object tracking by around 1400% to 1200%. The results presented in this work indicate that the NCS2 can achieve more than 50 FPS (frames per second) in image recognition and more than 20 FPS in object tracking. The power efficiency obtained by using the NCS2 can vary from 200% to 400%. These findings highlight the significant performance gains NCS2 offers in constrained hardware environments.