Smart Head-mount Obstacle Avoidance Wearable for the Vision Impaired
Blind and visually impaired (BVI) individuals face numerous challenges in navigation, including detecting obstacles, interpreting ground information, recognizing path boundaries, and obtaining spatial and directional data. Traditional tools like white canes are widely used due to their simplicity and effectiveness in detecting ground-level obstacles. However, these canes have limitations in providing information about higher-level obstacles or distant objects, which can lead to collisions and hinder independent mobility. Therefore, recognizing and avoiding obstacles, particularly those at head height or in crowded areas, remains a key challenge.
Technological advancements have led to the development of assistive devices aimed at enhancing navigation safety and independence. Many Electronic Travel Aids (ETAs) have emerged, utilizing radar, ultrasound, electromagnetic technology, image and other sensing technologies to assist users with safe navigation and obstacle avoidance. Despite the advancements, there are still significant gaps in the effectiveness and user adoption of ETAs. Most approaches have been product-centric or technology-centric, overlooking crucial aspects of user preferences and expectations. The challenges include optimizing system cost-efficiency, minimizing latency, enhancing sustainability, improving device compactness, and achieving high detection accuracy. Therefore, in this research, we focus on BVI users’ expectations to develop solutions that enable comprehensive environmental awareness with a real-time, low-cost and energy efficient detection mechanism within a compact and portable device.
Firstly, we propose a unique head-mount hardware system including an ultrasound array and an inertial measurement unit to address human behavioural issues such as erroneous readings and head turns. A dataset collected from real-world scenarios validates the accuracy of the hardware in detecting obstacles while accommodating head movements. We use multiple machine learning classifiers to verify the reliability of the collected data and confirm the effectiveness of the designed prototype. The investigation focuses on developing low-cost, real-time detection mechanisms for resource-constrained wearable devices. The results prove that the proposed smart and efficient wearable solution, as a viable option for BVI individuals, is feasible.
In addition, we investigate the actual energy consumption of the machine learning models to ensure long operational times with limited battery power on the proposed device. The energy efficiency research is conducted to identify models balancing accuracy, real-time performance, and energy consumption. While high-performance deep learning algorithms excel in complex tasks, their deployment on wearable devices struggles due to limited battery power and computing capabilities. It is found that even small deep learning models, despite their high performance, are not satisfactory due to the computational cost and latency.
We further study the challenge faced by portable devices in maintaining sustainable real-time performance due to limited computational resources and battery capacity. Leveraging genetic programming for feature construction, we develop more energy-efficient models that enhance real-time obstacle detection. More condensed features are found by genetic programming, ensuring cost-effectiveness and sustainability without compromising accuracy. These models are validated on actual portable boards and Field Programmable Gate Array (FPGA) simulations, and comparative analyses are conducted using various performance metrics, including hardware resource utilization. The results show that genetic programming constructed models significantly reduce energy consumption and inference time with minimal accuracy loss. Moreover, these models can achieve higher accuracy than benchmarks, offering flexibility in balancing energy usage and accuracy based on user needs. We also investigated FPGA-based spiking neural network hardware accelerators for obstacle avoidance. These implementations achieved significant reductions in both cost and power consumption compared to spiking neural network implementations using high-end CPUs or GPUs.
