This paper presents the T-RexNet approach to detect small moving objects in videos by using a deep neural network. T-RexNet combines the advantages of Single-Shot-Detectors with a specific feature-extraction network, thus overcoming the known shortcomings of Single-Shot-Detectors in detecting small objects. The deep convolutional neural network includes two parallel paths: the first path processes both the original picture, in gray-scale format, and differences between consecutive frames; in the second path, differences between a set of three consecutive frames is only handled. As compared with generic object detectors, the method limits the depth of the convolutional network to make it less sensible to high-level features and easier to train on small objects. The simple, Hardware-efficient architecture attains its highest accuracy in the presence of videos with static framing. Deploying our architecture on the NVIDIA Jetson Nano edge-device shows its suitability to embedded systems. To prove the effectiveness and general applicability of the approach, real-world tests assessed the method performances in different scenarios, namely, aerial surveillance with the WPAFB 2009 dataset, civilian surveillance using the Chinese University of Hong Kong (CUHK) Square dataset, and fast tennis-ball tracking, involving a custom dataset. Experimental results prove that T-RexNet is a valid, general solution to detect small moving objects, which outperforms in this task generic existing object-detection approaches. The method also compares favourably with application-specific approaches in terms of the accuracy vs. speed trade-off.

State of the Art solutions for image classification are nowadays mostly based on Convolutional Neural Networks (CNNs), With respect to running them for inference, training CNNs is a much more resource demanding job. For this reason, embedded devices designed to execute neural networks are normally not suitable to train the same networks. Hand image classification is becoming increasingly popular in several application areas, by empowering embedded systems with the ability to recognise hand gestures or grasp types. In this field, making it possible to retrain the hand image classifier directly on the embedded device would allow an easy customization of classification task according to application specific needs. In this paper we present a method based on transfer learning which solves this problem by first extracting hand keypoints, using with a dedicated neural network, and then applying on the resulting data a lightweight classifier, easily retrainable directly on the target embedded system. The proposal has been validated on a new freely distributed dataset, outperforming state of the art solutions in the specific task.