Recognition Algorithm of Bad Driving Behavior in Surveillance Video Images Based on Deep Learning

In traffic monitoring, accurately identifying bad driving behavior is challenging due to dense traffic, numerous pedestrians, and a wide variety ofbehaviors. This study proposes a deep learning-based algorithm for recognizing such behaviors in surveillance video. To address the problems of insufficient segmentation accuracy in complex traffic scenes, frequent omission of small abnormal behaviors, and poor recognition robustness under multi-target interference, this paper proposes the following innovations: a segmentation strategy based on a maximum-threshold processing algorithm combined with particle swarm optimization, which segments surveillance video images by computing the total image entropy, extracts foreground human figures, and thereby improves target segmentation accuracy and anti-interference capability under complex backgrounds; A keyframe localization algorithm based on keypoint motion velocity and acceleration extracts a set of keyframe sequences containing the driver’s skeleton map and node sets from the foreground image, accurately segments the driving task, and overcomes the difficulty of capturing key features of small-scale bad behaviors; By feeding the keyframe sequence set into a spatio-temporal feature-enhanced graph convolutional network, the spatial and temporal feature vectors of driving behavior are extracted via the model’s graph convolution structure and hybrid pooling processing. Then, a deep embedding clustering algorithm in the output layer classifies the feature vectors, reducing the impact of diversity and variability of bad driving behaviors on recognition performance and yielding the final classification results for different categories in surveillance video images. Test results demonstrate high performance, with pixel segmentation accuracy, inter-region contrast, and intra-region consistency all exceeding 0.92, effectively separating individuals from their surroundings and improving the identification of various bad driving behaviors.The recognition robustness in complex traffic scenarios is significantly better than the comparison methods.