Using Deep Neural Networks for Lane Change Identification

Examensarbete för masterexamen
Widjaja, Ryan Damarputra
Lane change maneuvers are commonly performed by drivers in highway driving situations. It starts with the driver planning and making the decision whether a lane change maneuver is necessary according to the situation. Once the driver decides to do the maneuver, he/she starts to prepare themselves. Next, he/she changes their lateral position until the vehicle crosses the line between lanes. Finally after crossing the lane marking, the driver adapts to his/her new positions by stabilizing the vehicle. When lane change maneuvers are done incorrectly, accidents may happened which can be fatal to those involved in the crash. Advanced Driver Assistance Systems (ADAS) include several functions that rely on lane change detection e.g., lane departure warning (LDW) system and lane change assistance system. These functions can be used to help driver perform a safer lane change maneuver. Having a system which can accurately retrieve and recognize the driving characteristics of a lane change maneuver will be beneficial for the development of ADAS. Identifying lane changes in driving data can be done manually by annotations, but it costs a substantial amount of time and money in case of large driving databases. A cheaper solution would be to use machine learning algorithms as they excel in this type of problem. Several machine learning algorithms, specifically artificial neural networks, have been used in many different research applications, including lane change predictions. This thesis work included several steps. Firstly, driving data was retrieved from UDRIVE, which contains naturalistic driving data collected from various European countries. The data was processed into segments containing lane changes and baseline driving. It served as an input for training and testing using a sliding time window approach. Three neural networks were constructed to identify lane changes. One served as the baseline model, and the other two were variations of the baseline model called modified Long Short Term Memory (LSTM) and stacked LSTM, respectively. Training and testing were conducted to these networks using the same configuration and dataset. During the training process, some of the parameters were adjusted according to their performance and some could not be adjusted. Parameters that cannot be adjusted are called hyperparameters and they usually relate to the structure of a neural network model. Both the modified LSTM and stacked LSTM were subjected to parameter tuning, which is a process of changing various trainable parameter in order to make the model perform optimally. After parameter tuning was done, the best model was further evaluated by using cross validation. Results have shown that the stacked LSTM model has the best performance among the three models. It managed to reach F1 score of 0:7178 and able to identify 95% of the lane change data. However, the stacked LSTM model has performance problems in terms of training time in identifying the transition phases of lane change maneuver. Several factors which contributes to this performances issue are identified, such as the imbalanced training data, the variables selection, the structure of the network, the number of trainable parameters, the hyperparameter settings, and how the raw input data is processed. If these issues are resolved, it is expected that the stacked LSTM would have a higher performance in the range of 0:85 in F1 score.
lane change, machine learning, neural network, lstm
Arkitekt (konstruktör)
Geografisk plats
Byggnad (typ)
Teknik / material