At present, the sorting of agricultural products in China mainly relies on manual labour, which results in low efficiency, and the development of corresponding automatic equipment lags behind. The grasping method based on machine vision has been widely used in industry, and can provide a reference for the automatic sorting of agricultural products. In this paper, an automatic sorting device for agricultural products was designed. The grasping mechanism adopted a 4-degree-of-freedom (4-DOF) manipulator, and the machine vision control system adopted a monocular camera, which can realize the positioning and classification of the grasp-target. First, the geometric model of the manipulator was established, and the kinematics model of the manipulator was established via the Denavit-Hartenberg (D-H) parameter method. Next, the kinematics analysis and verification were carried out. Then, Zhang Zhengyou calibration method was used to calibrate the camera. An image processing method based on histogram correction was proposed. Based on this, a target positioning algorithm based on the pinhole imaging principle and a target classification algorithm based on the area threshold were established. Finally, an automatic sorting test platform for agricultural products using a visual servo was built. Target classification, positioning and sorting tests were conducted using tomatoes and oranges as the test objects. The test results show that the success rate of the target positioning is close to 98%, that of the target classification is close to 98% and that of the grasping is close to 95%. Furthermore, the sorting time of a single target object can be as fast as 1 second, which can meet the requirements of automatic sorting for common agricultural products. The automatic sorting device for agricultural products has a simple structure, reliable performance and low costs. The structure and algorithms proposed in this paper are simple, reliable, and highly efficient and thus can easily r...

The ability to make informed decisions in complex scenarios is crucial for intelligent automotive systems. Traditional expert rules and other methods often fall short in complex contexts. Recently, reinforcement learning has garnered significant attention due to its superior decision-making capabilities. However, there exists the phenomenon of inaccurate target network estimation, which limits its decision-making ability in complex scenarios. This paper mainly focuses on the study of the underestimation phenomenon, and proposes an end-to-end autonomous driving decision-making method based on an improved TD3 algorithm. This method employs a forward camera to capture data. By introducing a new critic network to form a triple-critic structure and combining it with the target maximization operation, the underestimation problem in the TD3 algorithm is solved. Subsequently, the multi-timestep averaging method is used to address the policy instability caused by the new single critic. In addition, this paper uses Carla platform to construct multi-vehicle unprotected left turn and congested lane-center driving scenarios and verifies the algorithm. The results demonstrate that our method surpasses baseline DDPG and TD3 algorithms in aspects such as convergence speed, estimation accuracy, and policy stability.

Predicting future trajectories is crucial for autonomous vehicles, as accurate predictions enhance safety and inform subsequent decision-making and planning modules. This is however a challenging task due to the complex interactions between surrounding vehicles. Existing methods struggled to extract deep representations and often overlook spatial dependence. To address this problem, this paper introduces GIVA, an interaction-aware trajectory prediction method based on the Gated Recurrent Unit (GRU)-Improved Visual Geometry Group (VGG)-Attention Mechanism model. GIVA first encodes the historical trajectories of the target vehicle and its surrounding vehicles using a GRU Encoder. Next, an Interaction Module, which combines the Improved VGG Pooling Module and the Attention Mechanism Pooling Module, effectively captures spatial interaction features between vehicles. The Improved VGG Pooling Module extracts more detailed and effective interaction information, while the Attention Mechanism Pooling Module emphasizes the importance of surrounding vehicles for the target vehicle’s future trajectory. Lastly, the dynamic encoding feature of the target vehicle and the fused interaction feature are concatenated and input into a GRU Decoder to generate the future trajectory. Experiments on the public Next Generation Simulation (NGSIM) dataset showcase the effectiveness of GIVA compared to existing prediction approaches, demonstrating its potential for improving autonomous vehicle performance.

Accurately predicting the future trajectory of surrounding vehicles is critical for autonomous driving. However, this task still faces challenges in reducing computational costs, modeling complex interactions, and enhancing prediction realism. To this end, we proposed a multimodal trajectory prediction model for autonomous vehicles based on VectorNet-Interaction-GAN, named VIG. Specifically, VIG takes the historical trajectories and high-definition map information as inputs. Then, the improved VectorNet is employed as a feature extraction module to encode the vectorized inputs, effectively avoiding the high rendering loss and computational density associated with rasterized methods. Next, to capture complex interactions, we introduce an interaction module comprising four sub-networks (Vehicles to Lanes, Lanes to Lanes, Lanes to Vehicles and Vehicles to Vehicles), which utilize the attention mechanism and the polyline subgraph to model the different entities. Finally, a GAN-based framework is incorporated to enhance the realism and plausibility of the predicted trajectories. Extensive experiments on the Argoverse motion dataset demonstrate that VIG outperforms the existing methods while maintaining low model complexity.

Accurate and efficient trajectory prediction of surrounding vehicles is critical for ensuring the safety of autonomous driving systems, as it provides vital input for subsequent decision-making and planning modules. Relying solely on historical trajectory information is insufficient to accurately predict, especially in complex traffic scenarios with interwoven roads such as ramps, intersections, and roundabouts. Therefore, predicting future trajectories requires consideration not only of the historical trajectories of vehicles but also of the interaction among various entities of the environmental context. However, existing prediction methods suffer from expensive rendering costs and entity information loss, unnecessary complexity and redundant computation, and all-to-all interaction modeling and high computational costs, facing the challenge of the trade-off of efficiency and accuracy. To this end, we propose the SGC-Informer-based deep learning framework for fast and accurate trajectory prediction, called FESIG. Specifically, a comprehensive explicit entity vectorized representation is constructed based on Feature Engineering for the agent's trajectories and High-Definition (HD) map information, avoiding expensive rendering costs and entity information loss. Then, to mitigate excess complexity, we introduce the Simple Graph Convolution (SGC) network to effectively extract the local feature of each entity, which can remove the nonlinear transformation of the information propagation. Additionally, the Informer mechanism is employed to automatically sort and select the entities in the scenario, and then capture the complex interactions between the target vehicle and these entities that significantly impact it, thereby reducing computation costs. Experiments performed on the Argoverse forecasting dataset illustrate that the FESIG achieves the trade-off between accuracy and efficiency in the vehicle trajectory prediction task compared with the existing methods.

With the rapid development of intelligent sensing and control technologies, map building and path planning for indoor mobile robots have been widely applied in indoor navigation scenarios. However, existing SLAM methods suffer from low loop closure detection accuracy and map drift, while existing global path planning methods often yield non-smooth paths with suboptimal search efficiency. To this end, we propose an integrated SLAM and path planning method leveraging an improved Gmapping algorithm and A* algorithm. Firstly, a SLAM method integrating the Gmapping-Cartographer algorithm is introduced, which incorporates an improved Gmapping algorithm into Cartographer’s loop closure detection module. This approach ensures mapping accuracy while effectively reducing mapping time and mitigating localization errors and map drift. Secondly, a path planning method based on the improved A* algorithm is introduced. The heuristic function of the A* algorithm is refined by adjusting the weights of the actual cost and the estimated cost, which accelerates search convergence and improves planning accuracy. Moreover, to alleviate the prevalence of excessive path inflection points and large turning angles, the Bézier curve method is introduced for smoothing, ensuring path fluidity and fewer turns. Finally, the effectiveness of the proposed method is validated in a Gazebo virtual simulation environment and a real laboratory environment. Experimental results demonstrate that our SLAM method demonstrates superior performance in mapping accuracy and successfully constructs high-precision maps. Simultaneously, our path planning method shows significant performance advantages over the original A*, ant colony optimization, and genetic algorithm in maps of different scales and achieves trajectories with fewer inflection points and greater smoothness.