A low-power real-time hidden Markov model accelerator for gesture user interface on wearable devices

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

“…The finger trajectory recognition system realizes real-time dynamic gesture recognition utilizing the hidden Markov model (HMM) [18]. The hand trajectory is divided into different strokes with different orientations before being input into the HMM model.…”

“…In addition to the aforementioned HGR systems [16]- [18] based on computing- intensive algorithms, some compact HGR systems [19], [20] adopt customized algorithms to achieve a balance between accuracy and power. However, the energyefficient features used in these systems limit the number of gesture types that can be recognized, therefore, constraining the applicable scenarios of the HGR systems.…”

“…However, the sophisticated control logic for improving the reconfigurability of the processing element array results in massive power consumption (630 mW) and large chip area (9.8 × 8.4mm2), which makes it less competitive for edge devices. The hand trajectory recognition system based on the hidden Markov model (HMM) [18] provides appealing power consumption (6.4 mW) through a binary search technique and a cluttered database. However, the computation-intensive tasks, such as image processing and fingertip localization, were implemented on FPGA whose power was not included in the power analysis.…”

“…To the best of our knowledge, the existing vision-based HGR systems recognize gestures based on the "one-shot" scheme, namely, classifying a gesture based on the appearance of one image. Therefore, to achieve high recognition accuracy, features in an image need to be fully extracted, leading to high computing complexity and power overhead [16]- [18]. On the other hand, energy-efficient HGR systems [14], [19], [20] based on simplified features usually undergo stability degradation, particularly when deployed under backgrounds with more interferences.…”

“…However, the complexity of the algorithms deployed in [16]- [18] is much higher than the proposed one, which results in massive power consumption. The vision chip in [17] Table 3- Nevertheless, the high flexibility comes with sophisticated control logic, which results in considerably high power overhead.…”