Adaptive Window Size Based Deep Neural Network for Driving Maneuver Prediction

Driving maneuver prediction is one of the most challenging tasks in Advanced Driver Assistance System (ADAS), it can provide an early notification for ADAS to predict dangerous circumstances and take appropriate actions. However, it is difficult to well modeling the driving maneuver process due to the complexity and uncertainty of traffic status. To address this issue, we propose a novel model, denoted as DMPM, which uses deep learning method for Driving Maneuver Prediction (DMP) from multi-modal data, i.e. front view videos and vehicle signals. Firstly, with Adaptive Window Size Selector (AWSS), DMPM is able to dynamically identify the optimal sliding window size for the input data. Secondly, the Global Context Video Network (GCV Network) is proposed combing with Root-ResNet+Weighted Channel Dropout (WCD) architecture to extract the features from multi-modal data efficiently. Specially, including the Global Context (GC) block, GCV Network has an ability of modeling long-range dependency. Finally, a Long Short-Term Memory (LSTM) network that captures temporal dependencies is leveraged for driving maneuver prediction. The experimental results show that the DMPM is capable of learning the latent features of driving maneuver and achieving significantly better performance than other popular models on a real-world driving data set.