T-RexNet-A Hardware-Aware Neural Network for Real-Time Detection of Small Moving Objects.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Single-Shot-Detectors struggle with detecting small objects.
• Existing generic object detectors are often less effective for small object detection.
• Need for efficient and accurate small object detection in surveillance and tracking.

Purpose of the Study:

• Introduce T-RexNet, a novel deep neural network for small moving object detection.
• Overcome the limitations of traditional Single-Shot-Detectors in identifying small objects.
• Evaluate T-RexNet's performance across diverse real-world scenarios.

Main Methods:

• Developed T-RexNet, a deep convolutional neural network with two parallel processing paths.
• The network processes grayscale images and frame differences for enhanced feature extraction.
• Limited network depth to improve sensitivity to small object features and ease of training.

Main Results:

• T-RexNet demonstrates superior accuracy in detecting small moving objects compared to generic detectors.
• The architecture achieves high accuracy, particularly in videos with static framing.
• Achieved a favorable accuracy-speed trade-off against application-specific methods.

Conclusions:

• T-RexNet provides a valid and generalizable solution for detecting small moving objects in videos.
• The approach is suitable for embedded systems, as demonstrated on the NVIDIA Jetson Nano.
• T-RexNet outperforms existing methods in small object detection accuracy and efficiency.