In response to the complex and diverse nature of the road traffic environment, where vehicle and pedestrian detection is prone to false and missed detections, this paper proposes a vehicle and pedestrian target detection algorithm YOLOv8-RC based on multi-scale feature fusion. Initially, the RCS-OSA module is introduced within the structure of the base network YOLOv8 to replace the original module, thereby enhancing and integrating the extracted feature information. Additionally, a lightweight Context-Aware Adaptive Feature Reorganization (CARAFE) is employed to replace the original upsampling operator, enhancing the network’s capability for global multi-scale information fusion. Subsequently, a detection dataset consisting of 6 000 images of vehicle and pedestrian targets is constructed through public datasets and network collection. The algorithm’s detection performance is quantitatively evaluated using accuracy, recall rate, mean Average Precision at a 50% intersection over union threshold (mAP50), and mAP50-95. Compared to YOLOv8-N, YOLOv8-RC demonstrates an improvement of 1.7 percentage in accuracy, 1.2 percentage in recall rate, 0.9 percentage in mAP50, and 0.5 percentage in mAP50-95, thus validating the algorithm’s effectiveness.

To address the issue of false and missed detection of non-motorized vehicles due to the small size and obstructed vision in autonomous vehicle target detection, this research refines YOLOv4 basic algorithm to bolster the accuracy of non-motorized vehicle detection. The optimized algorithm streamlines the feature extraction process through a cross-stage connection, concurrently diminishing computational overhead and bolstering detection efficiency. Additionally, Convolutional Block Attention Module (CBAM) is embedded to increase effective feature weights and improve detection accuracy through channel and spatial attention weights. A non-motorized vehicle detection model is established based on anchor adaptive matching using a self-built non-motorized vehicle dataset. To verify the effectiveness of the model, the performance of the model is compared through ablation experiments. The results show that the proposed detection model substantially improves the detection and recognition performance of non-motor vehicles, effectively solve the problems of missed and false detections.

To eliminate the defects of incomplete detection and high false detection rate caused by insignificant pedestrian target features, dense small targets and complex background in infrared images, this paper proposes an infrared pedestrian target detection algorithm based on improved YOLOv7. Firstly, the original Spatial Pyramid Pooling (SPP) module is replaced by the Channel Attention based Spatial Pyramid Pooling (CASPP) module based on the YOLOv7-tiny model, so that the model could pay more attention to the extraction of pedestrian features; then, the convolution module CBM based on the Meta-ACON activation function is introduced, which further suppressed the background noise and preserved the details of the pedestrians; finally, an alpha fusion data enhancement method is proposed to enrich the diversity of samples and improve the stability of the model in complex environments. The validation based on the FLIR dataset shows that the proposed method improves the accuracy by 3% and reduces the computation by 38% compared with the YOLOv7-tiny algorithm, which is more suitable for infrared pedestrian target detection scenarios.

In response to the multilane lines driving environments such as lane line occlusion, road shadows, where the extractes lane line feature information is missing, resulting in blurry and discontinuous predicted lane lines, this paper proposes a lightweight U²-Net network for lane line detection algorithm. Firstly, the Residual U-blocks (RSU) module of the lightweight U²-Net network and multi feature scale fusion are used to obtain globally informative lane line features; secondly, pixel-by-pixel threshold judgment is performed on lane line features, and the least square method is selected combined with the lane line cluster of Region Of Interest (ROI) to fit lane line, to achieve multilane line detection and determine the self lane line area; finally, the proposed lane detection algorithm is validated and analyzed in the TuSimple dataset. The results show that the average accuracy of the proposed lane line detection algorithm reaches 98.4%. Compared with other lane line detection networks, this algorithm has fewer network parameters and higher accuracy.

To satisfy the requirement of low power consumption vehicle computing platform for lane detection, this paper proposes a low computing power dependent real-time lane recognition method. Considering the variation of illumination during vehicle driving, a color separation method based on adaptive illumination to extract lane characteristics is proposed. The effective edge point form is defined and the lane lines are determined by edge point voting based on the classical edge detection and Hough transform algorithm. The lane lines are used to filter and supplement the edge points and the lane curve equation is obtained by using the random sample consensus algorithm. The results show that the proposed method achieves a recognition accuracy of over 98% and computation speed of 38 frames per second on a low power processor. Furthermore, the method has proven to be stable and robust in a variety of scenarios.

To address the issue of extensive blind spots during right turns due to the oversized nature of dump trucks, this paper proposes a dynamic detection algorithm for risk targets in the right-turn blind spots of dump trucks. The algorithm improves the YOLOv8 model by enhancing the C2f module and lossing calculation module to refine the model’s detection accuracy. Additionally, four position threshold lines are preset in the blind spots, the risk warning module of the blind spots of the dump truck is added, and the auxiliary driving system of the blind spots of the dump truck is established. The results indicate that the proposed dynamic detection algorithm can recognize various types of targets, including cars, trucks, buses, pedestrians and electric bicycles, with a mean Average Precision (mAP50) of 0.87 at a 50% intersection over union threshold for all categories of targets. The right-turn blind spots assisted driving system of the dump truck can make different degrees of early warning according to the position of the risk target box in the image.

To address the issue of large amount of computation, slow arithmetic speed of the current defogging video model that can only process single image, this article proposed a defogging simplified model for vehicular video based on guide filtering. Firstly, the law of atmospheric attenuation index of light energy and physical model of haze degradation were introduced, then the defogging simplified model was obtained by correcting the attenuation law and the degradation model, and two parameters of the fog concentration factor and ambient air light were determined with the gray scale method and guided filtering method. Finally, the model was verified by computer simulation. The simulation results show that the defogging simplified model works effectively to defog for continuous video.