Object tracking is an essential aspect of environmental perception technology for autonomous vehicles. The existing object tracking algorithms can only be applied well to simple scenes. When the scenes become complex, the algorithms have poor tracking performance and insufficient robustness, and the problems of tracking drift and object loss are prone to occur. Therefore, a robust object tracking algorithm for autonomous vehicles in complex scenes is proposed. Firstly, we study the Siam-FC network and related algorithms, and analyze the problems that need to be addressed in object tracking. Secondly, the construction of a double-template Siamese network model based on multi-feature fusion is described, as is the use of the improved MobileNet V2 as the feature extraction backbone network, and the attention mechanism and template online update mechanism are introduced. Finally, relevant experiments were carried out based on public datasets and actual driving videos, with the aim of fully testing the tracking performance of the proposed algorithm on different objects in a variety of complex scenes. The results showed that, compared with other algorithms, the proposed algorithm had high tracking accuracy and speed, demonstrated stronger robustness and anti-interference abilities, and could still accurately track the object in real time without the introduction of complex structures. This algorithm can be effectively applied in intelligent vehicle driving assistance, and it will help to promote the further development and improvement of computer vision technology in the field of environmental perception.

The need for a vehicle to perceive information about the external environmental as an independent intelligent individual has grown with the progress of intelligent driving from primary driver assistance to high-level autonomous driving. The ability of a common independent sensing unit to sense the external environment is limited by the sensor’s own characteristics and algorithm level. Hence, a common independent sensing unit fails to obtain comprehensive sensing information independently under conditions such as rain, fog, and night. Accordingly, an extended network-based fusion target detection algorithm for millimeter-wave radar and vision fusion is proposed in this work by combining the complementary perceptual performance of in-vehicle sensing elements, cost effectiveness, and maturity of independent detection technologies. Feature-level fusion is first used in this work according to the analysis of technical routes of the millimeter-wave radar and vision fusion. Training and test evaluation of the algorithm are carried out on the nuScenes dataset and test data from a homemade data acquisition platform. An extended investigation on the RetinaNet one-stage target detection algorithm based on the VGG-16+FPN backbone detection network is then conducted in this work to introduce millimeter-wave radar images as auxiliary information for visual image target detection. We use two-channel radar and three-channel visual images as inputs of the fusion network. We also propose an extended VGG-16 network applicable to millimeter-wave radar and visual fusion and an extended feature pyramid network. Test results showed that the mAP of the proposed network improves by 2.9% and the small target accuracy is enhanced by 18.73% compared with those of the reference network for pure visual image target detection. This finding verified the detection capability and algorithmic feasibility of the proposed extended fusion target detection network for visually insensitive targets.