Lightweight Driver Behavior Identification Model with Sparse Learning on In-Vehicle CAN-BUS Sensor Data

Abstract

This study focuses on driver-behavior identification and its application to finding embedded solutions in a connected car environment. We present a lightweight, end-to-end deep-learning framework for performing driver-behavior identification using in-vehicle controller area network (CAN-BUS) sensor data. The proposed method outperforms the state-of-the-art driver-behavior profiling models. Particularly, it exhibits significantly reduced computations (i.e., reduced numbers both of floating-point operations and parameters), more efficient memory usage (compact model size), and less inference time. The proposed architecture features depth-wise convolution, along with augmented recurrent neural networks (long short-term memory or gated recurrent unit), for time-series classification. The minimum time-step length (window size) required in the proposed method is significantly lower than that required by recent algorithms. We compared our results with compressed versions of existing models by applying efficient channel pruning on several layers of current models. Furthermore, our network can adapt to new classes using sparse-learning techniques, that is, by freezing relatively strong nodes at the fully connected layer for the existing classes and improving the weaker nodes by retraining them using data regarding the new classes. We successfully deploy the proposed method in a container environment using NVIDIA Docker in an embedded system (Xavier, TX2, and Nano) and comprehensively evaluate it with regard to numerous performance metrics.

Transfer Learning With Time Series Data: A Systematic Mapping Study

“…With this technique, certain nodes within layers can be frozen, while others are retrained. Ullah and Kim [95] apply sparse learning for TL in driver behavior identification. They prevent the forgetting of important knowledge by freezing strong nodes and retrain weaker ones in the target domain.…”

“…For a time series of length l with d dimensions, this leads to a matrix of size d × l. After normalizing the features' value ranges, the matrix can be interpreted as an image. This method is applied in [60], [91], [95]. Several other works transform time series data into the time-frequency domain and use the spectrogram as a visual time series representation.…”

Demand-Driven Data Acquisition for Large Scale Fleets

“…2 Functionality to let several independent parties work with the system, providing each party potentially different data. 3 Data can be transferred from the vehicle to an external system and persisted there. 4 The external system continuously provides the received sensor data to other systems via streaming.…”

“…Besides the commercial demand identified in the insurance market, academia also has many active fields of research that depend on vehicle data. Among these fields are driver behavior identification [3], inference of lane change intentions [4], or drowsiness detection [5]. In addition, access to vehicle data is essential for municipalities to facilitate the transition to Smart Cities [6].…”

Intelligent Perception System of Robot Visual Servo for Complex Industrial Environment

“…For instance, Long et al [ 17 ] proposed a full convolutional neural network (FCNN) [ 17 ], where the fully connected layer in the common image classification network was replaced with a convolutional layer, and deconvolution was used to generate a segmented image of the same size as the original picture directly, realizing end-to-end image semantic segmentation. The SegNet convolutional network adopts a one-to-one corresponding encoder–decoder structure, where the encoder performs the maximum pooling operation and records the index position of the pooling, which is then used by the decoder to perform nonlinear sampling; by using this structure [ 18 ], the SegNet effectively improves segmentation accuracy [ 19 ]. The U-Net network can be used for the binary semantic segmentation of medical images, and its main advantage is a fewer number of model parameters, which makes it able to complete training on a small-scale dataset, achieving good results [ 20 ].…”