Infrared Image Object Detection of Vehicle and Person Based on Improved YOLOv5

Existing object detection algorithms are difficult to perform object detection tasks on embedded devices under the limitations of energy efficiency ratio and power consumption due to complex network structure and huge computational and parametric quantities. The object detection task in infrared images has low recognition rate and high false alarm rate due to long distance, weak energy and low resolution. In order to achieve the detection task at the mobile edge of infrared vehicle pedestrian target detection, this paper puts the YOLOv5 algorithm into a series of optimizations and proposes a lightweight YOLO-mini network structure. That is, instead of CSPDarknet, the MobileNetV2 network structure is used as the backbone feature extraction network with the addition of coordinate attention mechanism. Also, to make the network model more lightweight, the weights are converted to int8 type by quantized sensing training, which enables the task of the object detection algorithm for infrared vehicle pedestrian dataset on embedded devices. Experiments testing the FLIR dataset on NVIDIA Xavier NX show that this algorithm greatly reduces the number of network model parameters with less loss of accuracy and improves the FPS. mAP of YOLO-MobileNetV2 reaches 86.75%, number of parameters 2.76M, and FPS of 45; The network structure of YOLO-mini achieves 84.63% mAP, 0.69M number of parameters, and 63 FPS.