YOLO V5-MAX: A Multi-object Detection Algorithm in Complex Scenes

Abstract

The target detection of autonomous ground vehicles (AGVs) has the problem of few and slow object categories, which will cause great safety problems for AGVs. This paper proposes a YOLO v5-MAX algorithm to deal with the problem of a few types of object detection in complex scenes, e.g., city traffic jam, pedestrians crossing the road, running red lights, overtaking, merging, etc. The proposed algorithm consists of two parts. Firstly, the proposed algorithm uses YOLO v5s as the initial network model to train the vehicle detection model, which is used to detect the three categories of cars, buses, and trucks. Secondly, based on the first part, a Neck network and Head output layer are added to the proposed algorithm to detect four categories of person, bike, motor, and rider. In this paper, the most commonly used YOLO v5 object detection network is taken as an example to verify the effectiveness and realizability of our innovation. Of course, our method can also be applied to other object detection models, providing a theoretically feasible method for multi-object detection in the future. Finally, after the proposed algorithm is trained, it is deployed to Jetson TX2 for actual AGVs detection experiments. The experimental results show that the detection types and detection speed of the proposed algorithm have been greatly improved.