Fast and low-power behavior analysis on vehicles using smartphones

More and more deep learning methods are applied in unmanned or assisted driving, and have achieved very excellent performance. This paper describes long short-term memory recurrent neural networks used in assisted driving, which can capture the long temporal dependencies of multiple vehicles sensors' data, supporting drivers' behavior analysis on vehicles. Some optimization methods, such as model compression, weight quantization, adaptive window segmentation, are applied to make the deep network faster and less power. Therefore, it can be easily deployed on smart-phones and other embedded devices due to its moderate energy consumption and low latency. The architecture was trained in a sequence-to-sequence prediction manner, and it explicitly learns to predict the driving patterns given the temporal context. The experiment is executed on the smart-phone. Experimental results for different parameters are also presented in the paper. At last, we reduce the model size to 77 KB, the processing time to 4.27 ms, and the power overhead is 7.7 mW, the percentage of improved performance by our optimizations is over 60%.