Driver Activities Detection System Based on Optimized ReSNet101 Deep Learning Model and Transfer Learning

Recently, deep learning models -based classifiers workflow have been established a robust platform with assist of standard datasets to distinguish and clarify human's behaviours, including driver activities throughout driving.Driver activity detection systems can meaningfully upgrade road safety benchmarks and decrease the traffic accidents.Deep learning models have addressed activities detection problem through auto deep feature extraction layers, yet their detection accuracy requires massive and detailed samples for network training process.It's ordinarily arduous to improve a model using limited number of samples for gaining higher performance to accurately detecting the driver activity class due to the computation complexity of data collection and experiments that required for predicting the drivers activities/behaviours.In this paper, a proposed system of driver activities detection using improved structure of the ResNet101 deep learning model with transfer learning technique is presented.The major objective is to classify and localize six types of driver activities in order to decrease the consuming time of the computations, obtain higher detection accuracy of driver activities through achieving specific feature extraction.The proposed improvement methodology of ResNet101 model structure was achieved by adding custom layers for handling the grey-scale and colour images, activity classification, box boundary coordinate determination, and identity layer.Furthermore, we have adopted the LeckyRelu activation function and regression loss function for locating the desired activity accurately.An adjustment of the optimized model hyperparameters was achieved throughout training process.The proposed optimized model was evaluated based on variant experiments implemented on the Driver Inattention Detection dataset.The findings of model evaluation exhibited an average detection accuracy of 96.95% for all classes.