No abstract

In this paper, one solution for an end-to-end deep neural network for autonomous driving is presented. The main objective of our work was to achieve autonomous driving with a light deep neural network suitable for deployment on embedded automotive platforms. There are several end-to-end deep neural networks used for autonomous driving, where the input to the machine learning algorithm are camera images and the output is the steering angle prediction, but those convolutional neural networks are significantly more complex than the network architecture we are proposing. The network architecture, computational complexity, and performance evaluation during autonomous driving using our network are compared with two other convolutional neural networks that we re-implemented with the aim to have an objective evaluation of the proposed network. The trained model of the proposed network is four times smaller than the PilotNet model and about 250 times smaller than AlexNet model. While complexity and size of the novel network are reduced in comparison to other models, which leads to lower latency and higher frame rate during inference, our network maintained the performance, achieving successful autonomous driving with similar efficiency compared to autonomous driving using two other models. Moreover, the proposed deep neural network downsized the needs for real-time inference hardware in terms of computational power, cost, and size.

A new method for estimation of angles of leg segments and joints, which uses accelerometer arrays attached to body segments, is described. An array consists of two accelerometers mounted on a rigid rod. The absolute angle of each body segment was determined by band pass filtering of the differences between signals from parallel axes from two accelerometers mounted on the same rod. Joint angles were evaluated by subtracting absolute angles of the neighboring segments. This method eliminates the need for double integration as well as the drift typical for double integration. The efficiency of the algorithm is illustrated by experimental results involving healthy subjects who walked on a treadmill at various speeds, ranging between 0.15 m/s and 2.0 m/s. The validation was performed by comparing the estimated joint angles with the joint angles measured with flexible goniometers. The discrepancies were assessed by the differences between the two sets of data (obtained to be below 6 degrees) and by the Pearson correlation coefficient (greater than 0.97 for the knee angle and greater than 0.85 for the ankle angle).

Alternation of walking pattern decreases quality of life and may result in falls and injuries. Freezing of gait (FOG) in Parkinson's disease (PD) patients occurs occasionally and intermittently, appearing in a random, inexplicable manner. In order to detect typical disturbances during walking, we designed an expert system for automatic classification of various gait patterns. The proposed method is based on processing of data obtained from an inertial sensor mounted on shank. The algorithm separates normal from abnormal gait using Pearson's correlation and describes each stride by duration, shank displacement, and spectral components. A rule-based data processing classifies strides as normal, short (short(+)) or very short (short(-)) strides, FOG with tremor (FOG(+)) or FOG with complete motor block (FOG(-)). The algorithm also distinguishes between straight and turning strides. In 12 PD patients, FOG(+) and FOG(-) were identified correctly in 100% of strides, while normal strides were recognized in 95% of cases. Short(+) and short(-) strides were identified in about 84% and 78%. Turning strides were correctly identified in 88% of cases. The proposed method may be used as an expert system for detailed stride classification, providing warning for severe FOG episodes and near-fall situations.

No abstract