Framework comparison of neural networks for automated counting of vehicles and pedestrians

This paper presents a comparison of three neural network frameworks used to make volumetric counts in an automated and continuous way. In addition to cars, the application count pedestrians. Frameworks used are: SSD Mobilenet retrained, SSD Mobilenet pre-trained, and GoogLeNet pre-trained. The evaluation data set has a total duration of 60 minutes and comes from three different cameras. Images from the real deployment videos are included when training to enrich the detectable cases. Traditional detection models applied to vehicle counting systems usually provide high values for cars seen from the front. However, when the observer or camera is on the side, some models have lower detection and classification values. A new data set with fewer classes reach similar performance values as trained methods with default data sets. Results show that for the class cars, recall and precision values are 0.97 and 0.90 respectively in the best case, making use of a trained model by default, while for the class people the use of a re-trained model provides better results with precision and recall values of 1 and 0.82.