A 184-μW Error-Tolerant Real-Time Hand Gesture Recognition System With Hybrid Tiny Classifiers Utilizing Edge CNN

Design of low-power gesture recognition system and self-organizing map processors

“…The proposed segmentation scheme improves the system's immunity to the skin-color distractors by combining the RGB and the depth information. As a result, the system achieves an average accuracy of 90.7%, which is 12% higher than that utilizing the skin-color model only [32].…”

“…Since the skin color shows a better clustering effect in the YCbCr color space, the skin-color pixels are binarized following Eq. (4-1) as in [32].…”

“…Besides, the majority voting scheme is used to improve the error-tolerance of the system by ensembling the recognition results of several adjacent frames. Apart from the aforementioned algorithm optimization techniques, the proposed HGR system also adopts the following hardware optimizations for reducing power consumption and processing latency [32]. The remainder of this chapter is organized as follows.…”

“…objects in the background) in the image and reduces the computational load of the posterior stages. Prior arts segment the hand using either the skin-color model [32] or the depth-based thresholding technique [25]- [27]. However, the color-based scheme faces performance degradation when the background contains skin-color blobs, such as human faces.…”

“…These shortcomings are largely attributed to their reliance on computation-intensive algorithms, rendering them unsuitable for integration into wearable devices. Our previous low-power HGR system illustrated in chapter 3 achieves high energy efficiency [32]. However, it suffers from accuracy degradation when being used under complex backgrounds where there exist other skin-color objects.…”