Bus Passenger Load Prediction: Challenges from an Industrial Experience

Evaluating the Performance of Video-Based Automated Passenger Counting Systems in Real-World Conditions: A Comparative Study

“…Although the formula used to evaluate accuracy was not the formula specified in VDV recommendation 457-2, there was a discrepancy between the measured accuracy (57.74%) of the commercial APC and the accuracy claimed on the product’s datasheet (see Table 1 ). By comparing the two APC systems, it is evident that the estimation of the commercial APC system is susceptible to many variabilities; from the development of the low-cost system and the subsequent collection of ground-truth data, we could understand and postulate that the video-based APC systems can be affected by the following external factors: (i) the boarding or alighting of large group of passengers may cause over- or under-counting [ 16 , 28 , 40 , 49 ]; (ii) passengers who step off (alighted) the bus and immediately re-board could lead to duplicated counts [ 37 ]; (iii) when the bus reaches the full capacity, the system faces difficulty in accurately counting passengers due to the crowd around the door [ 36 ]; (iv) passengers lingering at the door during the entire trip can lead to inaccuracies [ 28 ]; (v) technical settings such as positioning [ 37 , 49 ], resolution quality [ 25 ], lens cleanliness and device calibration of the device can impact accuracy; and (vi) the possibility for the commercial APC system to miss some passengers because it is switched off or re-started at each terminus when the bus stays idle. Factors like lighting conditions have affected the results in some studies [ 25 , 27 , 28 , 37 ], though the results of our model were not significantly affected by different lighting conditions.…”

“…These authors have tested different ML and DL algorithms for passenger detection [ 27 , 36 , 37 ]. In addition, there have been attempts to enhance existing APC estimations: some authors have tested different algorithms for real-time passenger counting [ 38 , 39 , 40 , 41 , 42 ], while others have introduced methods that can be applied in a post-processing phase [ 43 , 44 ].…”

“…Although the literature includes the development of both algorithms and low-cost devices, i.e., cost-effective devices that utilize affordable price and computational resources for counting passengers, and although work has been performed on improving commercial APC estimations, there have been only few studies that have used real operational data [ 16 , 17 , 25 , 27 , 33 , 36 , 37 , 38 , 39 , 40 , 49 , 50 ]. Among these studies using real operational data, none have sought to evaluate commercial APC systems in real-world operational conditions and to compare them with low-cost APC systems.…”