Detection and Classification of Urban Actors Through TensorFlow with an Android Device

Abstract

With the advancement of Artificial Intelligence, now it is possible to perform Neural Networks for detection and classification inside a device with limited hardware. We compare the performance of Deep Neural Networks (DNN). Model selected is Single Shot Detector (SSD) by default and re-trained, and frameworks are TensorFlow mobile versus TensorFlow lite. Default model contains 80 different classes of objects and the second one is our re-trained model with only 6 different classes based on urban actors (car, bus, truck, bicycle, motorcycle, person). The main goal is to build an object tracker for urban mobility with an Android device. Results are based in two metrics for object detection: mean Average Precision (mAP) and log-average miss rate. And for classification we report two metrics: precision (PRE) and recall (REC). We report inference time as an additional metric which is strongly related to hardware used. TensorFlow mobile is much slower than TensorFlow lite in terms of time of inference. Finally, re-trained model allows the integration of new scenarios, improving the detection rate.