Implementación de un sistema para el conteo volumétrico de objetos mediante redes neuronales convolucionales

Abstract

Currently, traffic congestion has become a problem that affects the whole society, both drivers who spend a lot of time in huge rows of cars, as well as bystanders who are overwhelmed by the amount of cars they find it difficult to get from one sidewalk to the other. For this, different solutions have been implemented to reduce this problem from the placement of traffic lights at strategic points, also the increase of lanes, but these solutions do not completely dissipate the traffic congestion. Consequently, the search for better handling of traffic information has given rise to the need to benefit from a vehicular and pedestrian detection system. That is why the idea of implementing a system for counting volumetric objects through neural networks arises. This research describes the process of detection and counting of objects in real time which will be of the following classes: vehicles, motorcycles, bicycles and people, through the use of convolutional neural networks, from the acquisition and preparation of a set of images , which will be labeled according to each class to which they belong. We used a total of 2320 images taken in the city of Machala that include cars, motorcycles, bicycles and people, of which 90% was chosen for training and 10% for validation. For the labeling of the images, the LabelImg tool was used, which is a graphical image annotation tool and supports the YOLO format. In training the images, the Yolov3 neural network was used, which uses a more efficient object detection algorithm in real time than other networks. To evaluate our network, the use of evaluation metrics was established, for which we have tested the network with the data set that has been built. The tests show that the network has an efficient detection performance with mAP (average accuracy) of 65% with around 8000 iterations, with 416 x 416 image entries in an NVIDIA Geforce 940M. What took quite some time just for training the images around about 5 days per workout. For object counting we used the TensorFlow API, Object Counting with which we achieved efficient results with the previously trained neural network, we used the RTSP protocol to obtain real-time video, with a Dahua-1218 camera with resolution 1280x720 (720p) and 40 frames per second (fps), which was installed in the streets Junín between Guabo and Kléber Franco due to security situations and because said address is considered an area with an influx of cars, pedestrians, motorcycles, among others . Performance tests of the counting system were performed with different files of trained weights, with different image quantities, as well as variations in the iterations. The results of these tests yielded favorable results for the trained network with 2320 images with around 8000 iterations with respect to accuracy (0.52), reference value (0.53) and a mAP of 65.22%. Thus, an accurate counting system is available that detects most objects that appear in real time.