These parameters were
reportedly effective to represent a driver’s status at the yellow onset in other literature [21]. No lane changing was considered in the fact that it was rare according to our field observations. 4.1.2. ANN Model Outputs In general, there are two B-Raf assay possible time moments able to be used to tell an occurrence of RLR: the all-red onset and at the all-red end. If the all-red onset is used, a subject vehicle would be considered a red-light runner when it has not reached the stop line but cannot completely stop according to its distance to the stop line, speed, and maximal possible deceleration. If the all-red end is used, a subject vehicle is considered a red-light runner when it is still within the intersection when the all-red clearance expires. In the all-red-end-based method, two factors are considered relevant regarding the ANN outputs, the DTI and speed. Using the all-red onset would have to assume that driver becomes the red-light runner only when it cannot stop. However, this assumption is questionable because a slow driver may still want to take the RLR risk to cross the intersection or it may just be distracted and become a red-light runner. Therefore, in this paper, a vehicle’s status at the all-red clearance end was used to measure the RLR event.
Two types of outputs were used: (1) classifier: a vehicle was labeled as a run-light runner if it was still within the intersection
at the all-red clearance end, regardless where it is exactly; (2) the vehicle’s location and speed were observed within the intersection at the all-red end. These two output variants were evaluated, respectively, and compared later. 4.2. ANN Model Design 4.2.1. ANN Structure Since the driver behaviors during the yellow and all-red are independent from cycle to cycle (i.e., the drivers are not aware of other vehicles’ maneuvers in previous cycles), the feedforward neuron network was selected due to its memoryless property. In terms of the number of hidden neurons, two ANN structures were used and compared: standard feedforward neuron network with multiple hidden layers and the cascade-correlation (CC) neuron network which iteratively inserts new hidden neurons until the CC can achieve the desired MSE. 4.2.2. Training Algorithms A variant of the generic backpropagation algorithm is used Batimastat to train the ANNs, namely, QuickProp. The readers can refer to the literature [13] for more details. The weights were updated once with (9) after all the sample pairs were fed in. In other words, the standard backpropagation algorithm in this paper was offline and the weights were updated only once after each epoch. 4.3. Experiments Design Although the RLR event is a severe safety problem, its occurrence is relatively rare at most intersections in reality.