With the widespread popularization of 5G communication technology, traditional object detection systems can no longer meet the demands of various industries transitioning to intelligence. Consumers have higher demands for traditional object detection systems that previously relied on machine learning, and related manufacturing enterprises are also beginning to transition towards intelligence and low power consumption. Among these, how to combine image enhancement technology, object detection technology, and neural networks, and port them to portable embedded platforms has become a hot research topic for developers today. This project designs an object detection system based on a heterogeneous processor, which can efficiently and accurately identify target objects in specific environments with dense object populations.

This paper first presents the overall design of the object detection system. For hardware design, the multi-core heterogeneous ARM processor RK3399 was selected as the core processor. The CAM1320 module was chosen for the image acquisition module, and peripheral auxiliary circuits were built to complete the hardware design. The software section introduces convolutional neural networks. Currently, mainstream object detection algorithms extract feature information from image pixel matrices through convolution and pooling, and perform class prediction on the images in the fully connected layer. The network structure, bounding boxes, and loss function of YOLOv3, a typical representative algorithm, are analyzed. By comparing the performance indicators of YOLOv3 with other object detection algorithms, YOLOv3 was ultimately chosen as the basis for algorithm improvement in this paper.

Subsequently, the software part of the system's object detection is elaborated. First, an HSI-based image